fix(runtime): release tail-drop frames before the last child

bend2/comp.ts

@@ -1029,6 +1029,21 @@ function ctr_doms(ctr: Bend.Ctr, xs?: HTerm[]): HTerm[] {
   return ctr_live(ctr, xs).map(([, , A]) => A);
 }
 
+// The slot term_drop moves last, so a node frees before its spine: the
+// last box of the last field that holds the node's own datatype.
+function ctr_spine(k: Name, n: number): number {
+  const ctr = FL.book.ctrs[k];
+  if (!ctr || n > WIDE) {
+    return 255;
+  }
+  const fam = Bend.book_fam(FL.book, k);
+  const lays = lay_node(k).arms![k];
+  const i = ctr_doms(ctr).findLastIndex((F) =>
+    ty_holds(F, (u) => u?.$ === "ADT" && u.k === fam));
+  return i < 0 ? 255 : lays.slice(0, i).reduce((at, lay) =>
+    at + lay.ks.length, lays[i].ks.lastIndexOf("box"));
+}
+
 function ctr_flds(k: Name, xs: HTerm[]): HTerm[] {
   const ds = FL.book.ctrs[k] ? ctr_tail(FL.book.ctrs[k]) : [];
   return xs.filter((_, j) => !ds[j] || dom_live(ds[j]));
@@ -2852,8 +2867,8 @@ export function compile_book(book: Bend.Book): string {
     `{ ${s.params.length}, ${s.frame === null ? 0
       : s.params.length - s.frame.at.length}, ${Number(FL.bangs.has(s.def))
       | Number(!forky.has(s.fid)) << 1} }`).join(", ")} };`,
-  `CONSTV u8 CID_T[][2] = { ${[...cids].map(([k, n]) =>
-    `{ ${n}, ${Number(FL.hot.has(k))} }`).join(", ")} };`,
+  `CONSTV u8 CID_T[][3] = { ${[...cids].map(([k, n]) =>
+    `{ ${n}, ${Number(FL.hot.has(k))}, ${ctr_spine(k, n)} }`).join(", ")} };`,
   `#define STAT_LEN ${FL.img.length}`, "",
   `#define WL_RESW ${resw}`, `#define BANGS   ${FL.bangs.size}`, "",
   `#define WL_BANK Term ${ws.join(", ")};`, "",
@@ -3918,6 +3933,26 @@ FAR void term_drop(Env e, Term t) {
         u32 cls = tag == TAG_ARR ? 64 | blk_cls(t)
           : n > ${WIDE} ? 64 | (n - 240)
           : cls_fit(tag == TAG_TSK ? n + 2 : n);
+        u32 s = tag == TAG_CTR ? CID_T[aux][2] : n;
+        if (s + 1 < n) {
+          Term f = H[loc + s];
+          H[loc + s] = H[loc + n - 1];
+          H[loc + n - 1] = f;
+        }
+        if (cls < 64) {
+          while (n > 1 && term_triv(H[loc + n - 1])) {
+            n -= 1;
+          }
+        } else {
+          u32 m = 1u << (cls & 63);
+          while (m > 1 && term_triv(H[loc + m - 1])) {
+            m -= 1;
+          }
+          if (m < 256) {
+            n   = m;
+            cls &= 63;
+          }
+        }
         c0 = H[loc];
         H[loc] = cur;
         cur = loc | ((u64)n << 48) | ((u64)cls << 56);
@@ -3940,22 +3975,21 @@ FAR void term_drop(Env e, Term t) {
           j = (u32)H[loc + 1];
         }
       }
-      if (j < n) {
-        Term c = j == 0 ? c0 : H[loc + j];
+      Term c = j >= n ? 0 : j == 0 ? c0 : H[loc + j];
+      if (j + 1 >= n) {
+        cur = H[loc];
+        heap_free(e, cls, loc);
+      } else {
         if (arr && j > 0) {
           H[loc + 1] = j + 1;
         }
         if (!arr || i < 2) {
           cur += 1ull << 40;
         }
-        if (!term_triv(c)) {
-          t = c;
-          break;
-        }
-      } else {
-        u64 up = H[loc];
-        heap_free(e, cls, loc);
-        cur = up;
+      }
+      if (!term_triv(c)) {
+        t = c;
+        break;
       }
     }
   }

tests/rfc/drop_array_spine.bend

@@ -0,0 +1,77 @@
+# drop order through an Array: term_drop frees an Array's frame once only
+# trivial cells follow, so a chain whose spine runs through one cell frees
+# each Array before it drops the next link. ACon moves its kids behind its
+# item. A 4-cell Array with the child in cell 1 becomes a plain frame of 2;
+# a 512-cell one stays an Array frame, with the child last (512) and in
+# cell 255 (live end 256, one past a plain frame). Each round drops fresh
+# chains whole, then sums fresh ones; a cell freed early would read back
+# wrong from reused memory in a later round. Expect 3 rounds of 3 * 5050.
+import Base
+
+type AT is Type:
+  ANil{}
+  ACon{kids: Array<AT>, item: List<&2, U32>}
+
+def first(xs: List<&2, U32>) -> U32:
+  match xs:
+    case []:
+      0
+    case x <> rest:
+      x
+
+def nils(+d: Nat) -> Array<AT>:
+  match d:
+    case 0n:
+      ALeaf{ANil{}}
+    case 1n+p:
+      ANode{nils(p), nils(p)}
+
+# the child in the last cell, or with mid in the last cell of the low half
+def kids(+d: Nat, mid: Bool, t: AT) -> Array<AT>:
+  match d mid:
+    case 0n _:
+      ALeaf{t}
+    case 1n+p True{}:
+      ANode{kids(p, False{}, t), nils(p)}
+    case 1n+p False{}:
+      ANode{nils(p), kids(p, False{}, t)}
+
+def at(+d: Nat, +mid: Bool, +n: Nat) -> AT:
+  match n:
+    case 0n:
+      ANil{}
+    case 1n+m:
+      ACon{kids(d, mid, at(d, mid, m)), [U32.from_nat(n)]}
+
+def pick(a: Array<AT>, mid: Bool) -> AT:
+  match a mid:
+    case ALeaf{t} _:
+      t
+    case ANode{xs, ys} True{}:
+      pick(xs, False{})
+    case ANode{xs, ys} False{}:
+      pick(ys, False{})
+
+def at_sum(+n: Nat, +mid: Bool, t: AT) -> U32:
+  match n t:
+    case 0n _:
+      0
+    case _ ANil{}:
+      0
+    case 1n+m ACon{ks, item}:
+      U32.add(first(item), at_sum(m, mid, pick(ks, mid)))
+
+def waste(a: AT, b: AT, c: AT, x: U32) -> U32:
+  x
+
+def round(+n: Nat) -> U32:
+  a = waste(at(2n, True{}, n), at(9n, False{}, n), at(9n, True{}, n), 0)
+  b = U32.add(at_sum(n, True{}, at(2n, True{}, n)), at_sum(n, False{}, at(9n, False{}, n)))
+  U32.add(a, U32.add(b, at_sum(n, True{}, at(9n, True{}, n))))
+
+def main() -> IO(Unit):
+  IO.print(Nat.show(U32.to_nat(round(100n)))
+    ++ " " ++ Nat.show(U32.to_nat(round(100n)))
+    ++ " " ++ Nat.show(U32.to_nat(round(100n))))
+
+#|15150 15150 15150

tests/rfc/drop_spine.bend

@@ -0,0 +1,108 @@
+# drop order: term_drop moves a constructor's recursive field to its
+# last slot and frees a frame once only trivial slots follow, so the
+# frame frees before the spine drops. RCon has the spine first, MCon
+# before a word, Rose behind a List and OCon behind a flat Maybe's tag;
+# each item is a boxed one-cell List. Each round drops a fresh owned copy
+# whole, then shares one, sums it through one holder and drops it through
+# the other; a misplaced field would read back wrong from reused memory
+# in a later round. Expect 3 rounds of 4 * 5050 + 100 * 7.
+import Base
+
+type RL is Data:
+  RNil{}
+  RCon{rest: RL, item: List<&2, U32>}
+
+type ML is Data:
+  MNil{}
+  MCon{item: List<&2, U32>, rest: ML, tag: U32}
+
+type OL is Data:
+  ONil{}
+  OCon{next: Maybe<&2, OL>, item: List<&2, U32>}
+
+type Rose is Data:
+  Rose{kids: List<&2, Rose>, item: List<&2, U32>}
+
+def rl(+n: Nat) -> RL:
+  match n:
+    case 0n:
+      RNil{}
+    case 1n+m:
+      RCon{rl(m), [U32.from_nat(n)]}
+
+def ml(+n: Nat) -> ML:
+  match n:
+    case 0n:
+      MNil{}
+    case 1n+m:
+      MCon{[U32.from_nat(n)], ml(m), 7}
+
+def ol(+n: Nat) -> OL:
+  match n:
+    case 0n:
+      ONil{}
+    case 1n+m:
+      OCon{Some{ol(m)}, [U32.from_nat(n)]}
+
+def rose(+n: Nat) -> Rose:
+  match n:
+    case 0n:
+      Rose{[], [0]}
+    case 1n+m:
+      Rose{[rose(m)], [U32.from_nat(n)]}
+
+def first(xs: List<&2, U32>) -> U32:
+  match xs:
+    case []:
+      0
+    case x <> rest:
+      x
+
+def rl_sum(xs: RL) -> U32:
+  match xs:
+    case RNil{}:
+      0
+    case RCon{rest, item}:
+      U32.add(first(item), rl_sum(rest))
+
+def ml_sum(xs: ML) -> U32:
+  match xs:
+    case MNil{}:
+      0
+    case MCon{item, rest, tag}:
+      U32.add(U32.add(first(item), tag), ml_sum(rest))
+
+def ol_sum(xs: OL) -> U32:
+  match xs:
+    case ONil{}:
+      0
+    case OCon{None{}, item}:
+      first(item)
+    case OCon{Some{rest}, item}:
+      U32.add(first(item), ol_sum(rest))
+
+def rose_sum(r: Rose) -> U32:
+  match r:
+    case Rose{[], item}:
+      first(item)
+    case Rose{k <> rest, item}:
+      U32.add(first(item), rose_sum(k))
+
+def waste(r: RL, m: ML, o: Rose, l: OL, x: U32) -> U32:
+  x
+
+def round(+n: Nat) -> U32:
+  a = waste(rl(n), ml(n), rose(n), ol(n), 0)
+  +r = rl(n)
+  +m = ml(n)
+  +o = rose(n)
+  +l = ol(n)
+  b = U32.add(U32.add(rl_sum(r), ml_sum(m)), U32.add(rose_sum(o), ol_sum(l)))
+  U32.add(a, waste(r, m, o, l, b))
+
+def main() -> IO(Unit):
+  IO.print(Nat.show(U32.to_nat(round(100n)))
+    ++ " " ++ Nat.show(U32.to_nat(round(100n)))
+    ++ " " ++ Nat.show(U32.to_nat(round(100n))))
+
+#|20900 20900 20900