diff --git a/lang/src/lib.rs b/lang/src/lib.rs
index 59995f11..0de1fd84 100644
--- a/lang/src/lib.rs
+++ b/lang/src/lib.rs
@@ -25,6 +25,7 @@
     pointer_is_aligned_to,
     maybe_uninit_fill
 )]
+#![feature(array_chunks)]
 #![warn(clippy::dbg_macro)]
 #![expect(internal_features)]
 #![no_std]
diff --git a/lang/src/son.rs b/lang/src/son.rs
index a52d3802..a7c02aa1 100644
--- a/lang/src/son.rs
+++ b/lang/src/son.rs
@@ -124,24 +124,26 @@ impl Default for Nodes {
 }
 
 impl Nodes {
-    fn loop_depth(&self, target: Nid) -> LoopDepth {
+    fn loop_depth(&self, target: Nid, scheds: Option<&[Nid]>) -> LoopDepth {
         self[target].loop_depth.set(match self[target].kind {
             Kind::Region | Kind::Entry | Kind::Then | Kind::Else | Kind::Call { .. } | Kind::If => {
                 if self[target].loop_depth.get() != 0 {
                     return self[target].loop_depth.get();
                 }
-                self.loop_depth(self[target].inputs[0])
+                self.loop_depth(self[target].inputs[0], scheds)
             }
             Kind::Loop => {
-                if self[target].loop_depth.get() == self.loop_depth(self[target].inputs[0]) + 1 {
+                if self[target].loop_depth.get()
+                    == self.loop_depth(self[target].inputs[0], scheds) + 1
+                {
                     return self[target].loop_depth.get();
                 }
-                let depth = self.loop_depth(self[target].inputs[0]) + 1;
+                let depth = self.loop_depth(self[target].inputs[0], scheds) + 1;
                 self[target].loop_depth.set(depth);
                 let mut cursor = self[target].inputs[1];
                 while cursor != target {
                     self[cursor].loop_depth.set(depth);
-                    let next = self.idom(cursor);
+                    let next = self.idom(cursor, scheds);
                     debug_assert_ne!(next, 0);
                     if matches!(self[cursor].kind, Kind::Then | Kind::Else) {
                         debug_assert_eq!(self[next].kind, Kind::If);
@@ -159,17 +161,21 @@ impl Nodes {
         self[target].loop_depth.get()
     }
 
-    fn idepth(&self, target: Nid) -> IDomDepth {
+    fn idepth(&self, target: Nid, scheds: Option<&[Nid]>) -> IDomDepth {
         if target == VOID {
             return 0;
         }
         if self[target].depth.get() == 0 {
             let depth = match self[target].kind {
                 Kind::End | Kind::Start => unreachable!("{:?}", self[target].kind),
-                Kind::Region => {
-                    self.idepth(self[target].inputs[0]).max(self.idepth(self[target].inputs[1]))
+                Kind::Region => self
+                    .idepth(self[target].inputs[0], scheds)
+                    .max(self.idepth(self[target].inputs[1], scheds)),
+                _ if self[target].kind.is_pinned() => self.idepth(self[target].inputs[0], scheds),
+                _ if let Some(scheds) = scheds => {
+                    self.idepth(scheds[target as usize], Some(scheds))
                 }
-                _ => self.idepth(self[target].inputs[0]),
+                _ => self.idepth(self[target].inputs[0], scheds),
             } + 1;
             self[target].depth.set(depth);
         }
@@ -195,15 +201,14 @@ impl Nodes {
         }
     }
 
-    fn push_up_impl(&mut self, node: Nid, visited: &mut BitSet) {
+    fn push_up_impl(&self, node: Nid, visited: &mut BitSet, scheds: &mut [Nid]) {
         if !visited.set(node) {
             return;
         }
 
-        for i in 1..self[node].inputs.len() {
-            let inp = self[node].inputs[i];
+        for &inp in &self[node].inputs[1..] {
             if !self[inp].kind.is_pinned() {
-                self.push_up_impl(inp, visited);
+                self.push_up_impl(inp, visited, scheds);
             }
         }
 
@@ -213,32 +218,17 @@ impl Nodes {
 
         let mut deepest = self[node].inputs[0];
         for &inp in self[node].inputs[1..].iter() {
-            if self.idepth(inp) > self.idepth(deepest) {
+            if self.idepth(inp, Some(scheds)) > self.idepth(deepest, Some(scheds)) {
                 if self[inp].kind.is_call() {
                     deepest = inp;
                 } else {
                     debug_assert!(!self.is_cfg(inp));
-                    deepest = self.idom(inp);
+                    deepest = self.idom(inp, Some(scheds));
                 }
             }
         }
 
-        if deepest == self[node].inputs[0] {
-            return;
-        }
-
-        let current = self[node].inputs[0];
-
-        let index = self[current].outputs.iter().position(|&p| p == node).unwrap();
-        self[current].outputs.remove(index);
-        self[node].inputs[0] = deepest;
-        debug_assert!(
-            !self[deepest].outputs.contains(&node) || self[deepest].kind.is_call(),
-            "{node} {:?} {deepest} {:?}",
-            self[node],
-            self[deepest]
-        );
-        self[deepest].outputs.push(node);
+        scheds[node as usize] = deepest;
     }
 
     fn collect_rpo(&self, node: Nid, rpo: &mut Vec<Nid>, visited: &mut BitSet) {
@@ -253,21 +243,21 @@ impl Nodes {
         rpo.push(node);
     }
 
-    fn push_up(&mut self, rpo: &mut Vec<Nid>, visited: &mut BitSet) {
+    fn push_up(&self, rpo: &mut Vec<Nid>, visited: &mut BitSet, scheds: &mut [Nid]) {
         debug_assert!(rpo.is_empty());
         self.collect_rpo(VOID, rpo, visited);
 
         for &node in rpo.iter().rev() {
-            self.loop_depth(node);
+            self.loop_depth(node, Some(scheds));
             for i in 0..self[node].inputs.len() {
-                self.push_up_impl(self[node].inputs[i], visited);
+                self.push_up_impl(self[node].inputs[i], visited, scheds);
             }
 
             if matches!(self[node].kind, Kind::Loop | Kind::Region) {
                 for i in 0..self[node].outputs.len() {
                     let usage = self[node].outputs[i];
                     if self[usage].kind == Kind::Phi {
-                        self.push_up_impl(usage, visited);
+                        self.push_up_impl(usage, visited, scheds);
                     }
                 }
             }
@@ -290,14 +280,14 @@ impl Nodes {
         rpo.clear();
     }
 
-    fn better(&mut self, is: Nid, then: Nid) -> bool {
-        debug_assert_ne!(self.idepth(is), self.idepth(then), "{is} {then}");
-        self.loop_depth(is) < self.loop_depth(then)
-            || self.idepth(is) > self.idepth(then)
+    fn better(&self, is: Nid, then: Nid, scheds: Option<&[Nid]>) -> bool {
+        debug_assert_ne!(self.idepth(is, scheds), self.idepth(then, scheds), "{is} {then}");
+        self.loop_depth(is, scheds) < self.loop_depth(then, scheds)
+            || self.idepth(is, scheds) > self.idepth(then, scheds)
             || self[then].kind == Kind::If
     }
 
-    fn is_forward_edge(&mut self, usage: Nid, def: Nid) -> bool {
+    fn is_forward_edge(&self, usage: Nid, def: Nid) -> bool {
         match self[usage].kind {
             Kind::Phi => {
                 self[usage].inputs[2] != def || self[self[usage].inputs[0]].kind != Kind::Loop
@@ -307,20 +297,27 @@ impl Nodes {
         }
     }
 
-    fn push_down(&mut self, node: Nid, visited: &mut BitSet, antideps: &mut [Nid]) {
+    fn push_down(
+        &self,
+        node: Nid,
+        visited: &mut BitSet,
+        antideps: &mut [Nid],
+        scheds: &mut [Nid],
+        antidep_bounds: &mut Vec<Nid>,
+    ) {
         if !visited.set(node) {
             return;
         }
 
         for usage in self[node].outputs.clone() {
             if self.is_forward_edge(usage, node) && self[node].kind == Kind::Stre {
-                self.push_down(usage, visited, antideps);
+                self.push_down(usage, visited, antideps, scheds, antidep_bounds);
             }
         }
 
         for usage in self[node].outputs.clone() {
             if self.is_forward_edge(usage, node) {
-                self.push_down(usage, visited, antideps);
+                self.push_down(usage, visited, antideps, scheds, antidep_bounds);
             }
         }
 
@@ -331,56 +328,60 @@ impl Nodes {
         let mut min = None::<Nid>;
         for i in 0..self[node].outputs.len() {
             let usage = self[node].outputs[i];
-            let ub = self.use_block(node, usage);
-            min = min.map(|m| self.common_dom(ub, m)).or(Some(ub));
+            let ub = self.use_block(node, usage, Some(scheds));
+            min = min.map(|m| self.common_dom(ub, m, Some(scheds))).or(Some(ub));
         }
         let mut min = min.unwrap();
 
-        debug_assert!(self.dominates(self[node].inputs[0], min));
+        debug_assert!(self.dominates(scheds[node as usize], min, Some(scheds)));
 
         let mut cursor = min;
-        while cursor != self[node].inputs[0] {
-            cursor = self.idom(cursor);
-            if self.better(cursor, min) {
+        let mut fuel = self.values.len();
+        while cursor != scheds[node as usize] {
+            debug_assert!(fuel != 0);
+            fuel -= 1;
+            cursor = self.idom(cursor, Some(scheds));
+            if self.better(cursor, min, Some(scheds)) {
                 min = cursor;
             }
         }
 
         if self[node].kind == Kind::Load {
-            min = self.find_antideps(node, min, antideps);
-        }
-
-        if self[node].kind == Kind::Stre {
-            antideps[node as usize] = self[node].inputs[0];
+            min = self.find_antideps(node, min, antideps, scheds, antidep_bounds);
         }
 
         if self[min].kind.ends_basic_block() {
-            min = self.idom(min);
+            min = self.idom(min, Some(scheds));
         }
 
-        self.assert_dominance(node, min, true);
+        self.assert_dominance(node, min, true, Some(scheds));
 
-        let prev = self[node].inputs[0];
-        debug_assert!(self.idepth(min) >= self.idepth(prev));
-        let index = self[prev].outputs.iter().position(|&p| p == node).unwrap();
-        self[prev].outputs.remove(index);
-        self[node].inputs[0] = min;
-        self[min].outputs.push(node);
+        debug_assert!(
+            self.idepth(min, Some(scheds)) >= self.idepth(scheds[node as usize], Some(scheds))
+        );
+        scheds[node as usize] = min;
     }
 
-    fn find_antideps(&mut self, load: Nid, mut min: Nid, antideps: &mut [Nid]) -> Nid {
+    fn find_antideps(
+        &self,
+        load: Nid,
+        mut min: Nid,
+        antideps: &mut [Nid],
+        scheds: &[Nid],
+        antidep_bounds: &mut Vec<Nid>,
+    ) -> Nid {
         debug_assert!(self[load].kind == Kind::Load);
 
         let (aclass, _) = self.aclass_index(self[load].inputs[1]);
 
         let mut cursor = min;
-        while cursor != self[load].inputs[0] {
+        while cursor != scheds[load as usize] {
             antideps[cursor as usize] = load;
             if self[cursor].clobbers.get(aclass as _) {
                 min = self[cursor].inputs[0];
                 break;
             }
-            cursor = self.idom(cursor);
+            cursor = self.idom(cursor, Some(scheds));
         }
 
         if self[load].inputs[2] == MEM {
@@ -390,16 +391,19 @@ impl Nodes {
         for out in self[self[load].inputs[2]].outputs.clone() {
             match self[out].kind {
                 Kind::Stre => {
-                    let mut cursor = self[out].inputs[0];
-                    while cursor != antideps[out as usize] {
+                    let mut cursor = scheds[out as usize];
+                    while cursor != scheds[load as usize]
+                        && self.idepth(cursor, Some(scheds))
+                            > self.idepth(scheds[load as usize], Some(scheds))
+                    {
                         if antideps[cursor as usize] == load {
-                            min = self.common_dom(min, cursor);
+                            min = self.common_dom(min, cursor, Some(scheds));
                             if min == cursor {
-                                self.bind(load, out);
+                                antidep_bounds.extend([load, out]);
                             }
                             break;
                         }
-                        cursor = self.idom(cursor);
+                        cursor = self.idom(cursor, Some(scheds));
                     }
                     break;
                 }
@@ -409,12 +413,15 @@ impl Nodes {
                         .position(|&n| n == self[load].inputs[2])
                         .unwrap();
                     let mut cursor = self[self[out].inputs[0]].inputs[n];
-                    while cursor != antideps[out as usize] {
+                    while cursor != scheds[load as usize]
+                        && self.idepth(cursor, Some(scheds))
+                            > self.idepth(scheds[load as usize], Some(scheds))
+                    {
                         if antideps[cursor as usize] == load {
-                            min = self.common_dom(min, cursor);
+                            min = self.common_dom(min, cursor, Some(scheds));
                             break;
                         }
-                        cursor = self.idom(cursor);
+                        cursor = self.idom(cursor, Some(scheds));
                     }
                 }
                 _ => {}
@@ -431,9 +438,9 @@ impl Nodes {
         self[to].inputs.push(from);
     }
 
-    fn use_block(&self, target: Nid, from: Nid) -> Nid {
+    fn use_block(&self, target: Nid, from: Nid, scheds: Option<&[Nid]>) -> Nid {
         if self[from].kind != Kind::Phi {
-            return self.idom(from);
+            return self.idom(from, scheds);
         }
 
         let index = self[from].inputs.iter().position(|&n| n == target).unwrap_or_else(|| {
@@ -442,26 +449,28 @@ impl Nodes {
         self[self[from].inputs[0]].inputs[index - 1]
     }
 
-    fn idom(&self, target: Nid) -> Nid {
+    fn idom(&self, target: Nid, scheds: Option<&[Nid]>) -> Nid {
         match self[target].kind {
-            Kind::Start => VOID,
+            Kind::Start => unreachable!(),
             Kind::End => unreachable!(),
             Kind::Region => {
                 let &[lcfg, rcfg] = self[target].inputs.as_slice() else { unreachable!() };
-                self.common_dom(lcfg, rcfg)
+                self.common_dom(lcfg, rcfg, scheds)
             }
+            _ if self[target].kind.is_pinned() => self[target].inputs[0],
+            _ if let Some(scheds) = scheds => scheds[target as usize],
             _ => self[target].inputs[0],
         }
     }
 
-    fn common_dom(&self, mut a: Nid, mut b: Nid) -> Nid {
+    fn common_dom(&self, mut a: Nid, mut b: Nid, scheds: Option<&[Nid]>) -> Nid {
         while a != b {
-            let [ldepth, rdepth] = [self.idepth(a), self.idepth(b)];
+            let [ldepth, rdepth] = [self.idepth(a, scheds), self.idepth(b, scheds)];
             if ldepth >= rdepth {
-                a = self.idom(a);
+                a = self.idom(a, scheds);
             }
             if ldepth <= rdepth {
-                b = self.idom(b);
+                b = self.idom(b, scheds);
             }
         }
         a
@@ -581,16 +590,40 @@ impl Nodes {
         }
     }
 
-    fn gcm(&mut self, rpo: &mut Vec<Nid>, visited: &mut BitSet) {
+    fn gcm(&mut self, scratch: &mut Vec<Nid>, bind_buf: &mut Vec<Nid>, visited: &mut BitSet) {
         visited.clear(self.values.len());
-        self.fix_loops(rpo, visited);
+        self.fix_loops(bind_buf, visited);
+        debug_assert!(bind_buf.is_empty());
+        debug_assert!(scratch.is_empty());
+        scratch.resize(self.values.len() * 2, Nid::MAX);
+        let (antideps, scheds) = scratch.split_at_mut(self.values.len());
         visited.clear(self.values.len());
-        self.push_up(rpo, visited);
+        self.push_up(bind_buf, visited, scheds);
         visited.clear(self.values.len());
-        debug_assert!(rpo.is_empty());
-        rpo.resize(self.values.len(), VOID);
-        self.push_down(VOID, visited, rpo);
-        rpo.clear();
+        self.push_down(VOID, visited, antideps, scheds, bind_buf);
+
+        for &[from, to] in bind_buf.array_chunks() {
+            self.bind(from, to);
+        }
+
+        bind_buf.clear();
+        self[VOID].outputs =
+            self[VOID].outputs.iter().filter(|&&n| self[n].kind.is_at_start()).copied().collect();
+
+        for (&shed, n) in scheds.iter().zip(0u16..) {
+            if shed == Nid::MAX {
+                continue;
+            }
+
+            let prev = mem::replace(&mut self[n].inputs[0], shed);
+            if prev != VOID {
+                let index = self[prev].outputs.iter().position(|&o| o == n).unwrap();
+                self[prev].outputs.swap_remove(index);
+            }
+            self[shed].outputs.push(n);
+        }
+
+        scratch.clear();
     }
 
     fn clear(&mut self) {
@@ -1429,9 +1462,9 @@ impl Nodes {
                 }
                 return Some(self[target].inputs[2]);
             }
-            _ if self.is_cfg(target) && self.idom(target) == NEVER => panic!(),
-            K::Start
-            | K::Entry
+            K::Start => {}
+            _ if self.is_cfg(target) && self.idom(target, None) == NEVER => panic!(),
+            K::Entry
             | K::Mem
             | K::Loops
             | K::End
@@ -1468,7 +1501,7 @@ impl Nodes {
                 }
             }
 
-            cursor = self.idom(cursor);
+            cursor = self.idom(cursor, None);
         }
 
         CondOptRes::Unknown
@@ -1563,7 +1596,7 @@ impl Nodes {
     }
 
     #[expect(clippy::format_in_format_args)]
-    fn basic_blocks_instr(&mut self, out: &mut String, node: Nid) -> core::fmt::Result {
+    fn basic_blocks_instr(&self, out: &mut String, node: Nid) -> core::fmt::Result {
         match self[node].kind {
             Kind::Assert { .. } | Kind::Start => unreachable!("{} {out}", self[node].kind),
             Kind::End => return Ok(()),
@@ -1599,7 +1632,8 @@ impl Nodes {
         if self[node].kind != Kind::Loop && self[node].kind != Kind::Region {
             writeln!(
                 out,
-                " {:<14} {}",
+                " {:<3} {:<14} {}",
+                node,
                 format!("{:?}", self[node].inputs),
                 format!("{:?}", self[node].outputs)
             )?;
@@ -1609,7 +1643,7 @@ impl Nodes {
     }
 
     fn basic_blocks_low(
-        &mut self,
+        &self,
         out: &mut String,
         mut node: Nid,
         visited: &mut BitSet,
@@ -1709,7 +1743,7 @@ impl Nodes {
         Ok(())
     }
 
-    fn basic_blocks(&mut self) {
+    fn basic_blocks(&self) {
         let mut out = String::new();
         let mut visited = BitSet::default();
         self.basic_blocks_low(&mut out, VOID, &mut visited).unwrap();
@@ -1763,12 +1797,12 @@ impl Nodes {
             let mut stack = vec![self[loob].inputs[1]];
             let mut seen = BitSet::default();
             seen.set(loob);
-            let depth = self.loop_depth(loob);
+            let depth = self.loop_depth(loob, None);
             while let Some(nid) = stack.pop() {
                 if seen.set(nid) {
-                    if depth > self.loop_depth(nid) {
+                    if depth > self.loop_depth(nid, None) {
                         failed = true;
-                        log::error!("{depth} {} {nid} {:?}", self.loop_depth(nid), self[nid]);
+                        log::error!("{depth} {} {nid} {:?}", self.loop_depth(nid, None), self[nid]);
                     }
 
                     match self[nid].kind {
@@ -1830,25 +1864,25 @@ impl Nodes {
         aclass.last_store.set(lvar.last_store.get(), self);
     }
 
-    fn assert_dominance(&mut self, nd: Nid, min: Nid, check_outputs: bool) {
+    fn assert_dominance(&self, nd: Nid, min: Nid, check_outputs: bool, scheds: Option<&[Nid]>) {
         if !cfg!(debug_assertions) {
             return;
         }
 
         let node = self[nd].clone();
-        for &i in node.inputs.iter() {
-            let dom = self.idom(i);
+        for &i in &node.inputs[1..] {
+            let dom = self.idom(i, scheds);
             debug_assert!(
-                self.dominates(dom, min),
+                self.dominates(dom, min, scheds),
                 "{dom} {min} {node:?} {:?}",
                 self.basic_blocks()
             );
         }
         if check_outputs {
             for &o in node.outputs.iter() {
-                let dom = self.use_block(nd, o);
+                let dom = self.use_block(nd, o, scheds);
                 debug_assert!(
-                    self.dominates(min, dom),
+                    self.dominates(min, dom, scheds),
                     "{min} {dom} {node:?} {:?}",
                     self.basic_blocks()
                 );
@@ -1856,17 +1890,19 @@ impl Nodes {
         }
     }
 
-    fn dominates(&self, dominator: Nid, mut dominated: Nid) -> bool {
+    fn dominates(&self, dominator: Nid, mut dominated: Nid, scheds: Option<&[Nid]>) -> bool {
         loop {
             if dominator == dominated {
                 break true;
             }
 
-            if self.idepth(dominator) > self.idepth(dominated) {
+            debug_assert!(dominated != VOID);
+
+            if self.idepth(dominator, scheds) > self.idepth(dominated, scheds) {
                 break false;
             }
 
-            dominated = self.idom(dominated);
+            dominated = self.idom(dominated, scheds);
         }
     }
 
@@ -2018,8 +2054,11 @@ impl Kind {
     }
 
     fn is_pinned(&self) -> bool {
-        self.is_cfg()
-            || matches!(self, Self::Phi | Self::Arg | Self::Mem | Self::Loops | Kind::Assert { .. })
+        self.is_cfg() || self.is_at_start() || matches!(self, Self::Phi | Kind::Assert { .. })
+    }
+
+    fn is_at_start(&self) -> bool {
+        matches!(self, Self::Arg | Self::Mem | Self::Loops | Self::Entry)
     }
 
     fn is_cfg(&self) -> bool {
@@ -2384,7 +2423,8 @@ impl Ctx {
 pub struct Pool {
     cis: Vec<ItemCtx>,
     used_cis: usize,
-    nid_stack: Vec<Nid>,
+    scratch1: Vec<Nid>,
+    scratch2: Vec<Nid>,
     nid_set: BitSet,
 }
 
@@ -2652,8 +2692,8 @@ impl<'a> Codegen<'a> {
         if self.ci.nodes[value].kind == Kind::Load {
             let (lindex, ..) = self.ci.nodes.aclass_index(self.ci.nodes[value].inputs[1]);
             let clobber = self.ci.scope.aclasses[lindex].clobber.get();
-            if self.ci.nodes.idepth(clobber)
-                > self.ci.nodes.idepth(self.ci.scope.aclasses[index].clobber.get())
+            if self.ci.nodes.idepth(clobber, None)
+                > self.ci.nodes.idepth(self.ci.scope.aclasses[index].clobber.get(), None)
             {
                 self.ci.scope.aclasses[index].clobber.set(clobber, &mut self.ci.nodes);
             }
@@ -4732,7 +4772,7 @@ impl<'a> Codegen<'a> {
     fn finalize(&mut self, prev_err_len: usize) -> bool {
         use {AssertKind as AK, CondOptRes as CR};
 
-        self.ci.finalize(&mut self.pool.nid_stack, self.tys, self.files);
+        self.ci.finalize(&mut self.pool.scratch1, self.tys, self.files);
 
         //let mut to_remove = vec![];
         for (id, node) in self.ci.nodes.iter() {
@@ -4787,7 +4827,11 @@ impl<'a> Codegen<'a> {
         if self.errors.borrow().len() == prev_err_len {
             self.ci.nodes.check_final_integrity(self.ty_display(ty::Id::VOID));
             self.ci.nodes.graphviz(self.ty_display(ty::Id::VOID));
-            self.ci.nodes.gcm(&mut self.pool.nid_stack, &mut self.pool.nid_set);
+            self.ci.nodes.gcm(
+                &mut self.pool.scratch1,
+                &mut self.pool.scratch2,
+                &mut self.pool.nid_set,
+            );
             self.ci.nodes.check_loop_depth_integrity(self.ty_display(ty::Id::VOID));
             self.ci.nodes.basic_blocks();
             self.ci.nodes.graphviz(self.ty_display(ty::Id::VOID));
diff --git a/lang/src/son/hbvm.rs b/lang/src/son/hbvm.rs
index c8f10734..368d43c1 100644
--- a/lang/src/son/hbvm.rs
+++ b/lang/src/son/hbvm.rs
@@ -362,6 +362,9 @@ impl Nodes {
         let mut seen = BitSet::default();
         seen.clear(self.values.len());
 
+        let cfg_idx = outputs.iter().position(|&n| self.is_cfg(n)).unwrap();
+        outputs.swap(cfg_idx, 0);
+
         for &o in outputs.iter() {
             if (!self.is_cfg(o)
                 && self[o].outputs.iter().any(|&oi| {
@@ -407,6 +410,28 @@ impl Nodes {
             self[from]
         );
 
+        let bf = &buf;
+        debug_assert_eq!(
+            bf.iter()
+                .enumerate()
+                .filter(|(_, &b)| !self[b].kind.is_pinned())
+                .flat_map(|(i, &b)| self[b]
+                    .inputs
+                    .iter()
+                    .filter(|&&b| !self[b].kind.is_pinned())
+                    .filter_map(move |&inp| bf
+                        .iter()
+                        .position(|&n| inp == n)
+                        .filter(|&j| i > j)
+                        .map(|j| (bf[i], bf[j]))))
+                .collect::<Vec<_>>(),
+            vec![],
+            "{:?}",
+            bf
+        );
+
+        debug_assert!(self.is_cfg(bf[0]) || self[bf[0]].kind == Kind::Phi, "{:?}", self[bf[0]]);
+
         if outputs.len() != buf.len() {
             panic!("{:?} {:?}", outputs, buf);
         }
diff --git a/lang/src/son/hbvm/regalloc.rs b/lang/src/son/hbvm/regalloc.rs
index eac2e3ac..83143202 100644
--- a/lang/src/son/hbvm/regalloc.rs
+++ b/lang/src/son/hbvm/regalloc.rs
@@ -22,6 +22,7 @@ impl HbvmBackend {
         files: &[parser::Ast],
     ) -> (usize, bool) {
         let tail = Function::build(nodes, tys, &mut self.ralloc, sig);
+        nodes.basic_blocks();
 
         let strip_load = |value| match nodes[value].kind {
             Kind::Load { .. } if nodes[value].ty.loc(tys) == Loc::Stack => nodes[value].inputs[1],
@@ -366,6 +367,8 @@ impl<'a> Function<'a> {
         func.visited.clear(nodes.values.len());
         let mut s = Self { tail: true, nodes, tys, sig, func };
         s.emit_node(VOID);
+        debug_assert!(s.func.blocks.array_chunks().all(|[a, b]| a.end == b.start));
+        log::info!("{s:?}");
         s.tail
     }
 
@@ -513,9 +516,9 @@ impl Nodes {
     }
 
     fn use_block_of(&self, inst: Nid, uinst: Nid) -> Nid {
-        let mut block = self.use_block(inst, uinst);
+        let mut block = self.use_block(inst, uinst, None);
         while !self[block].kind.starts_basic_block() {
-            block = self.idom(block);
+            block = self.idom(block, None);
         }
         block
     }
@@ -538,7 +541,7 @@ impl Nodes {
 
     fn idom_of(&self, mut nid: Nid) -> Nid {
         while !self[nid].kind.starts_basic_block() {
-            nid = self.idom(nid);
+            nid = self.idom(nid, None);
         }
         nid
     }
@@ -643,12 +646,18 @@ impl<'a> Regalloc<'a> {
                 let mut range = b.range();
                 debug_assert!(range.start < range.end);
                 range.start = range.start.max(s.instr_of(inst).map_or(0, |n| n + 1) as usize);
-                debug_assert!(range.start < range.end, "{:?}", range);
+                debug_assert!(
+                    range.start < range.end,
+                    "{:?} {:?} {n} {inst}",
+                    range,
+                    self.nodes[inst]
+                );
                 let new = range.end.min(
                     s.instr_of(uinst)
                         .filter(|_| {
                             n == cursor
-                                && self.nodes.loop_depth(dom) == self.nodes.loop_depth(cursor)
+                                && self.nodes.loop_depth(dom, None)
+                                    == self.nodes.loop_depth(cursor, None)
                         })
                         .map_or(Nid::MAX, |n| n + 1) as usize,
                 );
@@ -700,10 +709,14 @@ impl<'a> Regalloc<'a> {
         debug_assert!(self.res.dfs_buf.is_empty());
         self.res.dfs_buf.push(from);
 
-        debug_assert!(self.nodes.dominates(until, from));
+        debug_assert!(self.nodes.dominates(until, from, None));
 
         while let Some(nid) = self.res.dfs_buf.pop() {
-            debug_assert!(self.nodes.dominates(until, nid), "{until} {:?}", self.nodes[until]);
+            debug_assert!(
+                self.nodes.dominates(until, nid, None),
+                "{until} {:?}",
+                self.nodes[until]
+            );
             each(self, nid, self.res.blocks[self.block_of(nid) as usize]);
             if nid == until {
                 continue;
diff --git a/lang/tests/son_tests_different_function_destinations.txt b/lang/tests/son_tests_different_function_destinations.txt
index adc56877..5919ee16 100644
--- a/lang/tests/son_tests_different_function_destinations.txt
+++ b/lang/tests/son_tests_different_function_destinations.txt
@@ -1,6 +1,6 @@
 main:
-    ADDI64 r254, r254, -144d
-    ST r31, r254, 80a, 64h
+    ADDI64 r254, r254, -136d
+    ST r31, r254, 80a, 56h
     LRA r32, r0, :glob_stru
     JAL r31, r0, :new_stru
     ST r1, r32, 0a, 16h
@@ -16,58 +16,57 @@ main:
     LI64 r32, 200d
     CP r1, r32
     JMP :1
- 2: LI64 r35, 3d
-    LI64 r36, 1d
-    ST r36, r32, 0a, 8h
-    ST r36, r32, 8a, 8h
-    ADDI64 r37, r254, 32d
-    ST r36, r254, 32a, 8h
-    ST r36, r254, 40a, 8h
-    ST r36, r254, 48a, 8h
-    ST r36, r254, 56a, 8h
-    ST r36, r254, 64a, 8h
-    ST r36, r254, 72a, 8h
+ 2: LI64 r34, 1d
+    ST r34, r32, 0a, 8h
+    ST r34, r32, 8a, 8h
+    ADDI64 r35, r254, 32d
+    ST r34, r254, 32a, 8h
+    ST r34, r254, 40a, 8h
+    ST r34, r254, 48a, 8h
+    ST r34, r254, 56a, 8h
+    ST r34, r254, 64a, 8h
+    ST r34, r254, 72a, 8h
+    LI64 r36, 3d
     CP r32, r33
- 8: JNE r32, r35, :3
+ 8: JNE r32, r36, :3
     LD r32, r254, 64a, 8h
     JEQ r32, r33, :4
     LI64 r32, 100d
     CP r1, r32
     JMP :1
- 4: ST r33, r254, 0a, 8h
+ 4: ST r34, r254, 32a, 8h
+    ST r34, r254, 40a, 8h
+    ST r34, r254, 48a, 8h
+    ST r34, r254, 56a, 8h
+    ST r34, r254, 64a, 8h
+    ST r34, r254, 72a, 8h
+    ST r33, r254, 0a, 8h
     ST r33, r254, 8a, 8h
     ST r33, r254, 16a, 8h
     ST r33, r254, 24a, 8h
-    ST r36, r254, 32a, 8h
-    ST r36, r254, 40a, 8h
-    ST r36, r254, 48a, 8h
-    ST r36, r254, 56a, 8h
-    ST r36, r254, 64a, 8h
-    ST r36, r254, 72a, 8h
     CP r32, r33
- 7: LD r38, r254, 64a, 8h
-    JNE r32, r35, :5
-    JEQ r38, r33, :6
+ 7: LD r37, r254, 64a, 8h
+    JNE r32, r36, :5
+    JEQ r37, r33, :6
     LI64 r32, 10d
     CP r1, r32
     JMP :1
  6: CP r1, r33
     JMP :1
- 5: ADD64 r34, r32, r36
-    MULI64 r32, r32, 16d
-    ADD64 r32, r37, r32
-    ST r33, r32, 0a, 8h
-    ST r33, r32, 8a, 8h
-    CP r32, r34
+ 5: MULI64 r37, r32, 16d
+    ADD64 r37, r35, r37
+    ST r33, r37, 0a, 8h
+    ST r33, r37, 8a, 8h
+    ADD64 r32, r32, r34
     JMP :7
- 3: MULI64 r34, r32, 16d
-    ADD64 r34, r37, r34
+ 3: MULI64 r37, r32, 16d
+    ADD64 r37, r35, r37
     JAL r31, r0, :new_stru
-    ST r1, r34, 0a, 16h
-    ADD64 r32, r32, r36
+    ST r1, r37, 0a, 16h
+    ADD64 r32, r32, r34
     JMP :8
- 1: LD r31, r254, 80a, 64h
-    ADDI64 r254, r254, 144d
+ 1: LD r31, r254, 80a, 56h
+    ADDI64 r254, r254, 136d
     JALA r0, r31, 0a
 new_stru:
     ADDI64 r254, r254, -16d
@@ -77,6 +76,6 @@ new_stru:
     LD r1, r13, 0a, 16h
     ADDI64 r254, r254, 16d
     JALA r0, r31, 0a
-code size: 739
+code size: 736
 ret: 0
 status: Ok(())
diff --git a/lang/tests/son_tests_generic_type_mishap.txt b/lang/tests/son_tests_generic_type_mishap.txt
index 4841b9f1..9028e5cf 100644
--- a/lang/tests/son_tests_generic_type_mishap.txt
+++ b/lang/tests/son_tests_generic_type_mishap.txt
@@ -10,9 +10,9 @@ opaque:
 process:
     ADDI64 r254, r254, -48d
     ST r31, r254, 16a, 32h
-    LI64 r32, 1000d
     ADDI64 r33, r254, 0d
     ST r0, r254, 0a, 1h
+    LI64 r32, 1000d
  4: JGTU r32, r0, :0
     JMP :1
  0: CP r2, r33
diff --git a/lang/tests/son_tests_generic_types.txt b/lang/tests/son_tests_generic_types.txt
index 4b6101d9..0c2f6733 100644
--- a/lang/tests/son_tests_generic_types.txt
+++ b/lang/tests/son_tests_generic_types.txt
@@ -69,8 +69,8 @@ new:
     ADDI64 r254, r254, 24d
     JALA r0, r31, 0a
 push:
-    ADDI64 r254, r254, -104d
-    ST r31, r254, 0a, 104h
+    ADDI64 r254, r254, -88d
+    ST r31, r254, 0a, 88h
     CP r38, r2
     CP r39, r3
     LI64 r37, 1d
@@ -95,12 +95,12 @@ push:
     LD r32, r38, 0a, 8h
     ADD64 r41, r32, r33
     CP r34, r35
- 7: LD r42, r38, 0a, 8h
-    LD r43, r38, 8a, 8h
+ 7: LD r33, r38, 0a, 8h
+    LD r36, r38, 8a, 8h
     JNE r41, r32, :5
-    JEQ r43, r0, :6
-    MUL64 r32, r43, r40
-    CP r2, r42
+    JEQ r36, r0, :6
+    MUL64 r32, r36, r40
+    CP r2, r33
     CP r3, r32
     CP r4, r40
     JAL r31, r0, :free
@@ -122,8 +122,8 @@ push:
     ADD64 r32, r32, r37
     ST r32, r38, 8a, 8h
     CP r1, r33
- 4: LD r31, r254, 0a, 104h
-    ADDI64 r254, r254, 104d
+ 4: LD r31, r254, 0a, 88h
+    ADDI64 r254, r254, 88d
     JALA r0, r31, 0a
 code size: 923
 ret: 69
diff --git a/lang/tests/son_tests_idk.txt b/lang/tests/son_tests_idk.txt
index a58d12f3..0f7074b6 100644
--- a/lang/tests/son_tests_idk.txt
+++ b/lang/tests/son_tests_idk.txt
@@ -1,17 +1,17 @@
 main:
     ADDI64 r254, r254, -128d
-    LI8 r15, 69b
-    LI64 r16, 128d
+    ADDI64 r15, r254, 0d
+    LI8 r16, 69b
+    LI64 r17, 128d
     CP r13, r0
-    ADDI64 r17, r254, 0d
- 2: LD r18, r254, 42a, 1h
-    JLTU r13, r16, :0
-    ANDI r13, r18, 255d
+ 2: LD r14, r254, 42a, 1h
+    JLTU r13, r17, :0
+    ANDI r13, r14, 255d
     CP r1, r13
     JMP :1
  0: ADDI64 r14, r13, 1d
-    ADD64 r13, r17, r13
-    ST r15, r13, 0a, 1h
+    ADD64 r13, r15, r13
+    ST r16, r13, 0a, 1h
     CP r13, r14
     JMP :2
  1: ADDI64 r254, r254, 128d
diff --git a/lang/tests/son_tests_inlining_issues.txt b/lang/tests/son_tests_inlining_issues.txt
index 906ddf04..b6fa85df 100644
--- a/lang/tests/son_tests_inlining_issues.txt
+++ b/lang/tests/son_tests_inlining_issues.txt
@@ -37,26 +37,26 @@ put_filled_rect:
     ST r7, r254, 75a, 1h
     ADDI64 r7, r254, 75d
     CP r16, r7
-    LI64 r17, 25d
-    LI64 r18, 2d
-    LI64 r19, 8d
-    ADDI64 r20, r254, 25d
-    ADDI64 r21, r254, 50d
-    LI8 r22, 5b
-    ST r22, r254, 25a, 1h
-    LD r23, r13, 0a, 8h
-    ST r23, r254, 26a, 4h
-    LI64 r24, 1d
-    ST r24, r254, 30a, 4h
+    ADDI64 r17, r254, 25d
+    LI8 r18, 5b
+    ST r18, r254, 25a, 1h
+    LD r19, r13, 0a, 8h
+    ST r19, r254, 26a, 4h
+    LI64 r20, 1d
+    ST r20, r254, 30a, 4h
     ST r16, r254, 34a, 8h
-    ST r22, r254, 50a, 1h
-    ST r23, r254, 51a, 4h
-    ST r24, r254, 55a, 4h
+    LI64 r21, 25d
+    ADDI64 r22, r254, 50d
+    ST r18, r254, 50a, 1h
+    ST r19, r254, 51a, 4h
+    ST r20, r254, 55a, 4h
     ST r16, r254, 59a, 8h
+    LI64 r23, 2d
+    LI64 r24, 8d
     LD r25, r15, 8a, 8h
     LD r13, r13, 8a, 8h
     ADD64 r26, r13, r25
-    SUB64 r26, r26, r24
+    SUB64 r26, r26, r20
     LD r27, r14, 8a, 8h
     MUL64 r26, r27, r26
     LD r14, r14, 0a, 8h
@@ -66,36 +66,36 @@ put_filled_rect:
     MUL64 r25, r27, r25
     ADD64 r14, r14, r25
     ADD64 r14, r28, r14
- 3: JGTU r13, r24, :0
-    JNE r13, r24, :1
+ 3: JGTU r13, r20, :0
+    JNE r13, r20, :1
     ADDI64 r13, r254, 0d
-    ST r22, r254, 0a, 1h
-    ST r23, r254, 1a, 4h
-    ST r24, r254, 5a, 4h
+    ST r18, r254, 0a, 1h
+    ST r19, r254, 1a, 4h
+    ST r20, r254, 5a, 4h
     ST r16, r254, 9a, 8h
     ST r14, r254, 17a, 8h
-    CP r2, r19
-    CP r3, r18
+    CP r2, r24
+    CP r3, r23
     CP r4, r13
-    CP r5, r17
+    CP r5, r21
     ECA 
     JMP :1
  1: JMP :2
  0: ST r14, r254, 67a, 8h
-    CP r2, r19
-    CP r3, r18
-    CP r4, r21
-    CP r5, r17
+    CP r2, r24
+    CP r3, r23
+    CP r4, r22
+    CP r5, r21
     ECA 
     ST r15, r254, 42a, 8h
-    CP r2, r19
-    CP r3, r18
-    CP r4, r20
-    CP r5, r17
+    CP r2, r24
+    CP r3, r23
+    CP r4, r17
+    CP r5, r21
     ECA 
+    SUB64 r13, r13, r23
     SUB64 r15, r15, r27
     ADD64 r14, r27, r14
-    SUB64 r13, r13, r18
     JMP :3
  2: ADDI64 r254, r254, 108d
     JALA r0, r31, 0a
diff --git a/lang/tests/son_tests_loops.txt b/lang/tests/son_tests_loops.txt
index 74632368..de4541c5 100644
--- a/lang/tests/son_tests_loops.txt
+++ b/lang/tests/son_tests_loops.txt
@@ -7,8 +7,8 @@ fib:
  2: JNE r13, r15, :0
     CP r1, r14
     JMP :1
- 0: ADD64 r14, r16, r14
-    SUB64 r13, r13, r17
+ 0: SUB64 r13, r13, r17
+    ADD64 r14, r16, r14
     SWA r14, r16
     JMP :2
  1: JALA r0, r31, 0a
diff --git a/lang/tests/son_tests_sort_something_viredly.txt b/lang/tests/son_tests_sort_something_viredly.txt
index 2351a62e..b05ab0b9 100644
--- a/lang/tests/son_tests_sort_something_viredly.txt
+++ b/lang/tests/son_tests_sort_something_viredly.txt
@@ -23,8 +23,8 @@ sqrt:
     ADD64 r18, r18, r15
     SLU64 r18, r18, r16
     JLTU r13, r18, :2
-    ADD64 r14, r15, r14
     SUB64 r13, r13, r18
+    ADD64 r14, r15, r14
     JMP :2
  2: SRUI64 r15, r15, 1b
     JMP :3
diff --git a/lang/tests/son_tests_string_flip.txt b/lang/tests/son_tests_string_flip.txt
index 351d7a3d..d737d87a 100644
--- a/lang/tests/son_tests_string_flip.txt
+++ b/lang/tests/son_tests_string_flip.txt
@@ -1,46 +1,45 @@
 main:
     ADDI64 r254, r254, -40d
     LI64 r17, 1d
-    LI64 r16, 4d
+    LI64 r15, 4d
     CP r14, r0
     ADDI64 r18, r254, 0d
     CP r13, r14
- 6: JNE r13, r16, :0
-    LI64 r19, 2d
-    ADDI64 r20, r254, 32d
+ 6: JNE r13, r15, :0
+    ADDI64 r19, r254, 32d
+    LI64 r20, 2d
     CP r13, r14
  4: LD r15, r254, 16a, 8h
     JNE r13, r17, :1
     CP r1, r15
     JMP :2
- 1: MUL64 r21, r13, r19
-    ADD64 r16, r13, r17
-    SUB64 r13, r19, r16
-    MUL64 r22, r13, r19
+ 1: ADD64 r16, r13, r17
+    SUB64 r15, r20, r16
+    MUL64 r21, r15, r20
+    MUL64 r22, r13, r20
     CP r13, r14
- 5: JNE r13, r19, :3
+ 5: JNE r13, r20, :3
     CP r13, r16
     JMP :4
  3: ADD64 r15, r13, r17
-    ADD64 r23, r21, r13
-    ADD64 r13, r22, r13
+    ADD64 r23, r22, r13
+    ADD64 r13, r21, r13
     MULI64 r23, r23, 8d
     MULI64 r13, r13, 8d
     ADD64 r23, r18, r23
     ADD64 r13, r18, r13
-    BMC r23, r20, 8h
+    BMC r23, r19, 8h
     BMC r13, r23, 8h
-    BMC r20, r13, 8h
+    BMC r19, r13, 8h
     CP r13, r15
     JMP :5
- 0: ADD64 r15, r13, r17
-    MULI64 r19, r13, 8d
-    ADD64 r19, r18, r19
-    ST r13, r19, 0a, 8h
-    CP r13, r15
+ 0: MULI64 r16, r13, 8d
+    ADD64 r16, r18, r16
+    ST r13, r16, 0a, 8h
+    ADD64 r13, r13, r17
     JMP :6
  2: ADDI64 r254, r254, 40d
     JALA r0, r31, 0a
-code size: 270
+code size: 267
 ret: 0
 status: Ok(())
diff --git a/lang/tests/son_tests_struct_patterns.txt b/lang/tests/son_tests_struct_patterns.txt
index 1a0cb139..55d641d3 100644
--- a/lang/tests/son_tests_struct_patterns.txt
+++ b/lang/tests/son_tests_struct_patterns.txt
@@ -28,8 +28,8 @@ fib_iter:
  2: JNE r13, r15, :0
     CP r1, r14
     JMP :1
- 0: ADD64 r14, r16, r14
-    SUB64 r13, r13, r17
+ 0: SUB64 r13, r13, r17
+    ADD64 r14, r16, r14
     SWA r14, r16
     JMP :2
  1: JALA r0, r31, 0a
diff --git a/lang/tests/son_tests_tests_ptr_to_ptr_copy.txt b/lang/tests/son_tests_tests_ptr_to_ptr_copy.txt
index 22ae15ad..18482953 100644
--- a/lang/tests/son_tests_tests_ptr_to_ptr_copy.txt
+++ b/lang/tests/son_tests_tests_ptr_to_ptr_copy.txt
@@ -8,9 +8,9 @@ main:
  4: JLTU r13, r17, :0
     LI64 r16, 10d
     CP r13, r14
- 3: LD r17, r254, 2048a, 1h
+ 3: LD r15, r254, 2048a, 1h
     JLTU r13, r16, :1
-    ANDI r13, r17, 255d
+    ANDI r13, r15, 255d
     CP r1, r13
     JMP :2
  1: ADD64 r15, r13, r14
diff --git a/lang/tests/son_tests_very_nested_loops.txt b/lang/tests/son_tests_very_nested_loops.txt
index 246a2557..3ba6fd74 100644
--- a/lang/tests/son_tests_very_nested_loops.txt
+++ b/lang/tests/son_tests_very_nested_loops.txt
@@ -1,24 +1,24 @@
 main:
-    LI64 r16, 4621819117588971520d
-    LI64 r17, 1d
+    LI64 r16, 1d
+    LI64 r17, 4621819117588971520d
     LI64 r18, 3d
     CP r14, r0
     CP r15, r14
  3: JNE r15, r18, :0
-    CP r1, r17
+    CP r1, r16
     JMP :1
  0: ITF64 r19, r15
     CP r13, r14
  5: JNE r13, r18, :2
-    ADD64 r15, r15, r17
+    ADD64 r15, r15, r16
     JMP :3
  2: ITF64 r20, r13
     FMUL64 r20, r20, r19
-    FCMPLT64 r20, r20, r16
+    FCMPLT64 r20, r20, r17
     NOT r20, r20
     ANDI r20, r20, 255d
     JNE r20, r0, :4
-    ADD64 r13, r13, r17
+    ADD64 r13, r13, r16
     JMP :5
  4: CP r1, r14
  1: JALA r0, r31, 0a
diff --git a/lang/tests/son_tests_wrong_dead_code_elimination.txt b/lang/tests/son_tests_wrong_dead_code_elimination.txt
index df3874ae..2f8bf099 100644
--- a/lang/tests/son_tests_wrong_dead_code_elimination.txt
+++ b/lang/tests/son_tests_wrong_dead_code_elimination.txt
@@ -1,8 +1,8 @@
 main:
     ADDI64 r254, r254, -1d
+    ST r0, r254, 0a, 1h
     LI64 r14, 255d
     LI8 r13, 1b
-    ST r0, r254, 0a, 1h
  2: LD r15, r254, 0a, 1h
     ANDI r16, r15, 255d
     CMPU r17, r16, r14