fixes #15076 (#16143)

* fixes #15076

* heapqueue: optimized for ARC

* added another test case [backport:1.4]

* code cleanup

4 files changed, 105 insertions, 8 deletions

diff --git a/compiler/injectdestructors.nim b/compiler/injectdestructors.nim
index 869e8ecc8..34c11e06c 100644
--- a/compiler/injectdestructors.nim
+++ b/compiler/injectdestructors.nim
@@ -37,7 +37,7 @@ type
     g: ControlFlowGraph
     graph: ModuleGraph
     otherRead: PNode
-    inLoop, inSpawn: int
+    inLoop, inSpawn, inLoopCond: int
     uninit: IntSet # set of uninit'ed vars
     uninitComputed: bool
     idgen: IdGenerator
@@ -296,8 +296,8 @@ proc isNoInit(dest: PNode): bool {.inline.} =
   result = dest.kind == nkSym and sfNoInit in dest.sym.flags
 
 proc genSink(c: var Con; dest, ri: PNode, isDecl = false): PNode =
-  if isUnpackedTuple(dest) or isDecl or
-      (isAnalysableFieldAccess(dest, c.owner) and isFirstWrite(dest, c)) or
+  if (c.inLoopCond == 0 and (isUnpackedTuple(dest) or isDecl or
+      (isAnalysableFieldAccess(dest, c.owner) and isFirstWrite(dest, c)))) or
       isNoInit(dest):
     # optimize sink call into a bitwise memcopy
     result = newTree(nkFastAsgn, dest, ri)
@@ -629,8 +629,10 @@ template handleNestedTempl(n, processCall: untyped, willProduceStmt = false) =
 
   of nkWhileStmt:
     inc c.inLoop
+    inc c.inLoopCond
     result = copyNode(n)
     result.add p(n[0], c, s, normal)
+    dec c.inLoopCond
     var bodyScope = nestedScope(s)
     let bodyResult = p(n[1], c, bodyScope, normal)
     result.add processScope(c, bodyScope, bodyResult)
diff --git a/lib/pure/collections/heapqueue.nim b/lib/pure/collections/heapqueue.nim
index 00e7a3b9d..30fa5dae8 100644
--- a/lib/pure/collections/heapqueue.nim
+++ b/lib/pure/collections/heapqueue.nim
@@ -65,7 +65,7 @@ proc len*[T](heap: HeapQueue[T]): int {.inline.} =
   ## Returns the number of elements of `heap`.
   heap.data.len
 
-proc `[]`*[T](heap: HeapQueue[T], i: Natural): T {.inline.} =
+proc `[]`*[T](heap: HeapQueue[T], i: Natural): lent T {.inline.} =
   ## Accesses the i-th element of `heap`.
   heap.data[i]
 
@@ -111,7 +111,7 @@ proc siftup[T](heap: var HeapQueue[T], p: int) =
   heap.data[pos] = newitem
   siftdown(heap, startpos, pos)
 
-proc push*[T](heap: var HeapQueue[T], item: T) =
+proc push*[T](heap: var HeapQueue[T], item: sink T) =
   ## Pushes `item` onto heap, maintaining the heap invariant.
   heap.data.add(item)
   siftdown(heap, 0, len(heap)-1)
@@ -168,7 +168,7 @@ proc del*[T](heap: var HeapQueue[T], index: Natural) =
   if index < newLen:
     heap.siftup(index)
 
-proc replace*[T](heap: var HeapQueue[T], item: T): T =
+proc replace*[T](heap: var HeapQueue[T], item: sink T): T =
   ## Pops and returns the current smallest value, and add the new item.
   ## This is more efficient than pop() followed by push(), and can be
   ## more appropriate when using a fixed-size heap. Note that the value
@@ -186,7 +186,7 @@ proc replace*[T](heap: var HeapQueue[T], item: T): T =
   heap.data[0] = item
   siftup(heap, 0)
 
-proc pushpop*[T](heap: var HeapQueue[T], item: T): T =
+proc pushpop*[T](heap: var HeapQueue[T], item: sink T): T =
   ## Fast version of a push followed by a pop.
   runnableExamples:
     var heap = initHeapQueue[int]()
@@ -197,7 +197,7 @@ proc pushpop*[T](heap: var HeapQueue[T], item: T): T =
     assert heap[0] == 6
     assert heap.pushpop(4) == 4
   result = item
-  if heap.len > 0 and heapCmp(heap.data[0], item):
+  if heap.len > 0 and heapCmp(heap.data[0], result):
     swap(result, heap.data[0])
     siftup(heap, 0)
 
diff --git a/tests/arc/tasyncleak4.nim b/tests/arc/tasyncleak4.nim
new file mode 100644
index 000000000..58cd7f0b7
--- /dev/null
+++ b/tests/arc/tasyncleak4.nim
@@ -0,0 +1,21 @@
+discard """
+  cmd: "nim c --gc:orc -d:useMalloc $file"
+  output: '''ok'''
+  valgrind: "leaks"
+"""
+
+# bug #15076
+import asyncdispatch
+
+var futures: seq[Future[void]]
+
+for i in 1..20:
+  futures.add sleepAsync 1
+  futures.add sleepAsync 1
+
+  futures.all.waitFor()
+  futures.setLen 0
+
+setGlobalDispatcher nil
+GC_fullCollect()
+echo "ok"
diff --git a/tests/arc/tdestroy_in_loopcond.nim b/tests/arc/tdestroy_in_loopcond.nim
new file mode 100644
index 000000000..62532664d
--- /dev/null
+++ b/tests/arc/tdestroy_in_loopcond.nim
@@ -0,0 +1,74 @@
+discard """
+  output: '''400 true'''
+  cmd: "nim c --gc:orc $file"
+"""
+
+type HeapQueue*[T] = object
+  data: seq[T]
+
+
+proc len*[T](heap: HeapQueue[T]): int {.inline.} =
+  heap.data.len
+
+proc `[]`*[T](heap: HeapQueue[T], i: Natural): T {.inline.} =
+  heap.data[i]
+
+proc push*[T](heap: var HeapQueue[T], item: T) =
+  heap.data.add(item)
+
+proc pop*[T](heap: var HeapQueue[T]): T =
+  result = heap.data.pop
+
+proc clear*[T](heap: var HeapQueue[T]) = heap.data.setLen 0
+
+
+type
+  Future = ref object of RootObj
+    s: string
+    callme: proc()
+
+var called = 0
+
+proc consume(f: Future) =
+  inc called
+
+proc newFuture(s: string): Future =
+  var r: Future
+  r = Future(s: s, callme: proc() =
+    consume r)
+  result = r
+
+var q: HeapQueue[tuple[finishAt: int64, fut: Future]]
+
+proc sleep(f: int64): Future =
+  q.push (finishAt: f, fut: newFuture("async-sleep"))
+
+proc processTimers =
+  # Pop the timers in the order in which they will expire (smaller `finishAt`).
+  var count = q.len
+  let t = high(int64)
+  while count > 0 and t >= q[0].finishAt:
+    q.pop().fut.callme()
+    dec count
+
+var futures: seq[Future]
+
+proc main =
+  for i in 1..200:
+    futures.add sleep(56020904056300)
+    futures.add sleep(56020804337500)
+    #futures.add sleep(2.0)
+
+    #futures.add sleep(4.0)
+
+    processTimers()
+    #q.pop()[1].callme()
+    #q.pop()[1].callme()
+
+    futures.setLen 0
+
+  q.clear()
+
+main()
+GC_fullCollect()
+echo called, " ", getOccupiedMem() < 160