GC: get rid of pathological behaviour for stack marking

lib/system.nim

@@ -2021,7 +2021,7 @@ proc `$`*[T: tuple|object](x: T): string =
   result = "("
   var firstElement = true
   for name, value in fieldPairs(x):
-    if not(firstElement): result.add(", ")
+    if not firstElement: result.add(", ")
     result.add(name)
     result.add(": ")
     result.add($value)

lib/system/gc.nim

@@ -1,7 +1,7 @@
 #
 #
 #            Nim's Runtime Library
-#        (c) Copyright 2013 Andreas Rumpf
+#        (c) Copyright 2015 Andreas Rumpf
 #
 #    See the file "copying.txt", included in this
 #    distribution, for details about the copyright.
@@ -1050,7 +1050,11 @@ when useMarkForDebug or useBackupGc:
     markGlobals(gch)
 
 proc collectCT(gch: var TGcHeap) =
-  if (gch.zct.len >= ZctThreshold or (cycleGC and
+  # stackMarkCosts prevents some pathological behaviour: Stack marking
+  # becomes more expensive with large stacks and large stacks mean that
+  # cells with RC=0 are more likely to be kept alive by the stack.
+  let stackMarkCosts = max(stackSize() div (16*sizeof(int)), ZctThreshold)
+  if (gch.zct.len >= stackMarkCosts or (cycleGC and
       getOccupiedMem(gch.region)>=gch.cycleThreshold) or alwaysGC) and
       gch.recGcLock == 0:
     when useMarkForDebug:

todo.txt

@@ -53,7 +53,6 @@ Bugs
 - VM: Pegs do not work at compile-time
 - VM: ptr/ref T cannot work in general
 - scopes are still broken for generic instantiation!
-- compilation of niminst takes way too long. looks like a regression
 - blocks can "export" an identifier but the CCG generates {} for them ...