Merge branch 'master' of github.com:Araq/Nimrod

Conflicts: lib/system/gc.nim
2026-01-05 20:47:53 +00:00 · 2013-02-11 07:11:42 +01:00
parent 9bb11a48a6 a668b4ffb5
commit 68135b3b50
24 changed files with 832 additions and 201 deletions
--- a/compiler/ccgstmts.nim
+++ b/compiler/ccgstmts.nim
@@ -1,7 +1,7 @@
 #
 #
 #           The Nimrod Compiler
-#        (c) Copyright 2012 Andreas Rumpf
+#        (c) Copyright 2013 Andreas Rumpf
 #
 #    See the file "copying.txt", included in this
 #    distribution, for details about the copyright.
@@ -15,6 +15,14 @@ const
  stringCaseThreshold = 8
    # above X strings a hash-switch for strings is generated

+proc registerGcRoot(p: BProc, v: PSym) =
+  if gSelectedGc == gcMarkAndSweep and containsGarbageCollectedRef(v.loc.t):
+    # we register a specialized marked proc here; this has the advantage
+    # that it works out of the box for thread local storage then :-)
+    let prc = genTraverseProcForGlobal(p.module, v)
+    linefmt(p.module.initProc, cpsStmts,
+      "#nimRegisterGlobalMarker($1);$n", prc)
+
 proc genVarTuple(p: BProc, n: PNode) = 
  var tup, field: TLoc
  if n.kind != nkVarTuple: InternalError(n.info, "genVarTuple")
@@ -28,6 +36,7 @@ proc genVarTuple(p: BProc, n: PNode) =
    if sfGlobal in v.flags:
      assignGlobalVar(p, v)
      genObjectInit(p, cpsInit, v.typ, v.loc, true)
+      registerGcRoot(p, v)
    else:
      assignLocalVar(p, v)
      initLocalVar(p, v, immediateAsgn=true)
@@ -143,7 +152,7 @@ proc genSingleVar(p: BProc, a: PNode) =
    # if sfImportc notin v.flags: constructLoc(p.module.preInitProc, v.loc)
    if sfExportc in v.flags and generatedHeader != nil:
      genVarPrototypeAux(generatedHeader, v)
-
+    registerGcRoot(p, v)
  else:
    assignLocalVar(p, v)
    initLocalVar(p, v, immediateAsgn)
--- a/compiler/ccgtrav.nim
+++ b/compiler/ccgtrav.nim
@@ -1,7 +1,7 @@
 #
 #
 #           The Nimrod Compiler
-#        (c) Copyright 2012 Andreas Rumpf
+#        (c) Copyright 2013 Andreas Rumpf
 #
 #    See the file "copying.txt", included in this
 #    distribution, for details about the copyright.
@@ -13,7 +13,7 @@
 # included from cgen.nim

 type
-  TTraversalClosure {.pure, final.} = object
+  TTraversalClosure = object
    p: BProc
    visitorFrmt: string

@@ -127,3 +127,20 @@ proc genTraverseProc(m: BModule, typ: PType, reason: TTypeInfoReason): PRope =
  m.s[cfsProcs].app(generatedProc)


+proc genTraverseProcForGlobal(m: BModule, s: PSym): PRope =
+  discard genTypeInfo(m, s.loc.t)
+  
+  var c: TTraversalClosure
+  var p = newProc(nil, m)
+  result = getGlobalTempName()
+  
+  c.visitorFrmt = "#nimGCvisit((void*)$1, 0);$n"
+  c.p = p
+  let header = ropef("N_NIMCALL(void, $1)()", result)
+  genTraverseProc(c, s.loc.r, s.loc.t)
+  
+  let generatedProc = ropef("$1 {$n$2$3$4}$n",
+        [header, p.s(cpsLocals), p.s(cpsInit), p.s(cpsStmts)])
+  
+  m.s[cfsProcHeaders].appf("$1;$n", header)
+  m.s[cfsProcs].app(generatedProc)
--- a/compiler/ccgtypes.nim
+++ b/compiler/ccgtypes.nim
@@ -1,7 +1,7 @@
 #
 #
 #           The Nimrod Compiler
-#        (c) Copyright 2012 Andreas Rumpf
+#        (c) Copyright 2013 Andreas Rumpf
 #
 #    See the file "copying.txt", included in this
 #    distribution, for details about the copyright.
@@ -938,7 +938,7 @@ proc genTypeInfo(m: BModule, typ: PType): PRope =
      genTupleInfo(m, fakeClosureType(t.owner), result)
  of tySequence, tyRef:
    genTypeInfoAux(m, t, result)
-    if usesNativeGC():
+    if gSelectedGC >= gcMarkAndSweep:
      let markerProc = genTraverseProc(m, t, tiNew)
      appf(m.s[cfsTypeInit3], "$1.marker = $2;$n", [result, markerProc])
  of tyPtr, tyRange: genTypeInfoAux(m, t, result)
--- a/compiler/commands.nim
+++ b/compiler/commands.nim
@@ -139,10 +139,11 @@ proc testCompileOptionArg*(switch, arg: string, info: TLineInfo): bool =
  case switch.normalize
  of "gc":
    case arg.normalize
-    of "boehm": result = gSelectedGC == gcBoehm
-    of "refc":  result = gSelectedGC == gcRefc
-    of "v2":    result = gSelectedGC == gcV2
-    of "none":  result = gSelectedGC == gcNone
+    of "boehm":        result = gSelectedGC == gcBoehm
+    of "refc":         result = gSelectedGC == gcRefc
+    of "v2":           result = gSelectedGC == gcV2
+    of "markandsweep": result = gSelectedGC == gcMarkAndSweep
+    of "none":         result = gSelectedGC == gcNone
    else: LocalError(info, errNoneBoehmRefcExpectedButXFound, arg)
  of "opt":
    case arg.normalize
@@ -276,6 +277,9 @@ proc processSwitch(switch, arg: string, pass: TCmdlinePass, info: TLineInfo) =
      gSelectedGC = gcRefc
    of "v2":
      gSelectedGC = gcV2
+    of "markandsweep":
+      gSelectedGC = gcMarkAndSweep
+      defineSymbol("gcmarkandsweep")
    of "none":
      gSelectedGC = gcNone
      defineSymbol("nogc")
--- a/compiler/ecmasgen.nim
+++ b/compiler/ecmasgen.nim
@@ -140,6 +140,8 @@ proc mangleName(s: PSym): PRope =
    app(result, toRope(s.id))
    s.loc.r = result

+proc makeJSString(s: string): PRope = strutils.escape(s).toRope
+
 proc genTypeInfo(p: var TProc, typ: PType): PRope
 proc genObjectFields(p: var TProc, typ: PType, n: PNode): PRope = 
  var 
@@ -165,7 +167,7 @@ proc genObjectFields(p: var TProc, typ: PType, n: PNode): PRope =
    s = genTypeInfo(p, field.typ)
    result = ropef("{kind: 1, offset: \"$1\", len: 0, " &
        "typ: $2, name: $3, sons: null}", 
-                   [mangleName(field), s, makeCString(field.name.s)])
+                   [mangleName(field), s, makeJSString(field.name.s)])
  of nkRecCase: 
    length = sonsLen(n)
    if (n.sons[0].kind != nkSym): InternalError(n.info, "genObjectFields")
@@ -193,7 +195,7 @@ proc genObjectFields(p: var TProc, typ: PType, n: PNode): PRope =
           [u, genObjectFields(p, typ, lastSon(b))])
    result = ropef("{kind: 3, offset: \"$1\", len: $3, " &
        "typ: $2, name: $4, sons: [$5]}", [mangleName(field), s, 
-        toRope(lengthOrd(field.typ)), makeCString(field.name.s), result])
+        toRope(lengthOrd(field.typ)), makeJSString(field.name.s), result])
  else: internalError(n.info, "genObjectFields")
  
 proc genObjectInfo(p: var TProc, typ: PType, name: PRope) = 
@@ -234,7 +236,7 @@ proc genEnumInfo(p: var TProc, typ: PType, name: PRope) =
    if i > 0: app(s, ", " & tnl)
    let extName = if field.ast == nil: field.name.s else: field.ast.strVal
    appf(s, "{kind: 1, offset: $1, typ: $2, name: $3, len: 0, sons: null}", 
-         [toRope(field.position), name, makeCString(extName)])
+         [toRope(field.position), name, makeJSString(extName)])
  var n = ropef("var NNI$1 = {kind: 2, offset: 0, typ: null, " &
      "name: null, len: $2, sons: [$3]};$n", [toRope(typ.id), toRope(length), s])
  s = ropef("var $1 = {size: 0, kind: $2, base: null, node: null, " &
@@ -588,7 +590,7 @@ proc genRaiseStmt(p: var TProc, n: PNode, r: var TCompRes) =
    typ = skipTypes(n.sons[0].typ, abstractPtrs)
    useMagic(p, "raiseException")
    appf(r.com, "raiseException($1, $2);$n", 
-         [a.res, makeCString(typ.sym.name.s)])
+         [a.res, makeJSString(typ.sym.name.s)])
  else: 
    useMagic(p, "reraiseException")
    app(r.com, "reraiseException();" & tnl)
@@ -626,7 +628,7 @@ proc genCaseStmt(p: var TProc, n: PNode, r: var TCompRes) =
          if stringSwitch: 
            case e.kind
            of nkStrLit..nkTripleStrLit: appf(r.com, "case $1: ", 
-                [makeCString(e.strVal)])
+                [makeJSString(e.strVal)])
            else: InternalError(e.info, "ecmasgen.genCaseStmt: 2")
          else: 
            appf(r.com, "case $1: ", [cond.res])
@@ -827,12 +829,12 @@ proc genFieldAddr(p: var TProc, n: PNode, r: var TCompRes) =
  var b = if n.kind == nkHiddenAddr: n.sons[0] else: n
  gen(p, b.sons[0], a)
  if skipTypes(b.sons[0].typ, abstractVarRange).kind == tyTuple:
-    r.res = makeCString("Field" & $getFieldPosition(b.sons[1]))
+    r.res = makeJSString("Field" & $getFieldPosition(b.sons[1]))
  else:
    if b.sons[1].kind != nkSym: InternalError(b.sons[1].info, "genFieldAddr")
    var f = b.sons[1].sym
    if f.loc.r == nil: f.loc.r = mangleName(f)
-    r.res = makeCString(ropeToStr(f.loc.r))
+    r.res = makeJSString(ropeToStr(f.loc.r))
  r.com = mergeExpr(a)

 proc genFieldAccess(p: var TProc, n: PNode, r: var TCompRes) = 
@@ -903,7 +905,7 @@ proc genAddr(p: var TProc, n: PNode, r: var TCompRes) =
        # globals are always indirect accessible
        r.kind = etyBaseIndex
        r.com = toRope("Globals")
-        r.res = makeCString(ropeToStr(s.loc.r))
+        r.res = makeJSString(ropeToStr(s.loc.r))
      elif sfAddrTaken in s.flags: 
        r.kind = etyBaseIndex
        r.com = s.loc.r
@@ -1422,8 +1424,8 @@ proc genReturnStmt(p: var TProc, n: PNode, r: var TCompRes) =
 proc genProcBody(p: var TProc, prc: PSym, r: TCompRes): PRope = 
  if optStackTrace in prc.options: 
    result = ropef("var F={procname:$1,prev:framePtr,filename:$2,line:0};$n" &
-        "framePtr = F;$n", [makeCString(prc.owner.name.s & '.' & prc.name.s), 
-                            makeCString(toFilename(prc.info))])
+        "framePtr = F;$n", [makeJSString(prc.owner.name.s & '.' & prc.name.s), 
+                            makeJSString(toFilename(prc.info))])
  else: 
    result = nil
  if p.beforeRetNeeded: 
@@ -1540,9 +1542,9 @@ proc gen(p: var TProc, n: PNode, r: var TCompRes) =
  of nkStrLit..nkTripleStrLit: 
    if skipTypes(n.typ, abstractVarRange).kind == tyString: 
      useMagic(p, "cstrToNimstr")
-      r.res = ropef("cstrToNimstr($1)", [makeCString(n.strVal)])
+      r.res = ropef("cstrToNimstr($1)", [makeJSString(n.strVal)])
    else: 
-      r.res = makeCString(n.strVal)
+      r.res = makeJSString(n.strVal)
  of nkFloatLit..nkFloat64Lit: 
    f = n.floatVal
    if f != f: r.res = toRope"NaN"
@@ -1611,8 +1613,8 @@ proc genModule(p: var TProc, n: PNode, r: var TCompRes) =
  if optStackTrace in p.options: 
    r.com = ropef("var F = {procname:$1,prev:framePtr,filename:$2,line:0};$n" &
        "framePtr = F;$n" & "$3" & "framePtr = framePtr.prev;$n", [
-        makeCString("module " & p.module.module.name.s), 
-        makeCString(toFilename(p.module.module.info)), r.com])
+        makeJSString("module " & p.module.module.name.s), 
+        makeJSString(toFilename(p.module.module.info)), r.com])

 proc myProcess(b: PPassContext, n: PNode): PNode = 
  if passes.skipCodegen(n): return n
--- a/compiler/nimrod.nim
+++ b/compiler/nimrod.nim
@@ -72,9 +72,9 @@ proc HandleCmdLine() =
 when defined(GC_setMaxPause):
  GC_setMaxPause 2_000

-when compileOption("gc", "v2"):
+when compileOption("gc", "v2") or compileOption("gc", "refc"):
  # the new correct mark&sweet collector is too slow :-/
  GC_disableMarkAndSweep()
 condsyms.InitDefines()
 HandleCmdLine()
-quit(options.gExitcode)
+quit(int8(msgs.gErrorCounter > 0))
--- a/compiler/options.nim
+++ b/compiler/options.nim
@@ -80,8 +80,8 @@ type                          # please make sure we have under 32 options
    cmdInteractive,           # start interactive session
    cmdRun                    # run the project via TCC backend
  TStringSeq* = seq[string]
-  TGCMode* = enum                   # the selected GC
-    gcNone, gcBoehm, gcRefc, gcV2   #
+  TGCMode* = enum             # the selected GC
+    gcNone, gcBoehm, gcMarkAndSweep, gcRefc, gcV2

 const
  ChecksOptions* = {optObjCheck, optFieldCheck, optRangeCheck, optNilCheck, 
@@ -269,7 +269,6 @@ proc binaryStrSearch*(x: openarray[string], y: string): int =
      return mid
  result = - 1

-# Can we keep this? I'm using it all the time
 template nimdbg*: expr = c.module.fileIdx == gProjectMainIdx
 template cnimdbg*: expr = p.module.module.fileIdx == gProjectMainIdx
 template pnimdbg*: expr = p.lex.fileIdx == gProjectMainIdx
--- a/doc/advopt.txt
+++ b/doc/advopt.txt
@@ -64,7 +64,8 @@ Advanced options:
  --skipUserCfg             do not read the user's configuration file
  --skipParentCfg           do not read the parent dirs' configuration files
  --skipProjCfg             do not read the project's configuration file
-  --gc:refc|v2|boehm|none   use Nimrod's native GC|V2|Boehm GC|no GC
+  --gc:refc|v2|markAndSweep|boehm|none
+                            select the GC to use; default is 'refc'
  --index:on|off            turn index file generation on|off
  --putenv:key=value        set an environment variable
  --babelPath:PATH          add a path for Babel support
--- a/doc/manual.txt
+++ b/doc/manual.txt
@@ -911,7 +911,21 @@ Arrays are always bounds checked (at compile-time or at runtime). These
 checks can be disabled via pragmas or invoking the compiler with the
 ``--boundChecks:off`` command line switch.

-The current implementation does not support nested open arrays.
+
+Open arrays
+-----------
+
+Often fixed size arrays turn out to be too inflexible; procedures should
+be able to deal with arrays of different sizes. The `openarray`:idx: type
+allows this; it can only be used for parameters. Openarrays are always 
+indexed with an ``int`` starting at position 0. The ``len``, ``low`` 
+and ``high`` operations are available for open arrays too. Any array with 
+a compatible base type can be passed to an openarray parameter, the index 
+type does not matter. In addition to arrays sequences can also be passed 
+to an open array parameter.
+
+The openarray type cannot be nested: multidimensional openarrays are not
+supported because this is seldom needed and cannot be done efficiently.


 Varargs
--- a/lib/system/cellsets.nim
+++ b/lib/system/cellsets.nim
@@ -1,7 +1,7 @@
 #
 #
 #            Nimrod's Runtime Library
-#        (c) Copyright 2012 Andreas Rumpf
+#        (c) Copyright 2013 Andreas Rumpf
 #
 #    See the file "copying.txt", included in this
 #    distribution, for details about the copyright.
@@ -39,7 +39,7 @@ type
    len, cap: int
    d: PCellArray

-# ------------------- cell set handling ---------------------------------------
+# ------------------- cell seq handling ---------------------------------------

 proc contains(s: TCellSeq, c: PCell): bool {.inline.} =
  for i in 0 .. s.len-1:
@@ -68,6 +68,8 @@ proc deinit(s: var TCellSeq) =
  s.len = 0
  s.cap = 0

+# ------------------- cell set handling ---------------------------------------
+
 const
  InitCellSetSize = 1024 # must be a power of two!

@@ -196,3 +198,21 @@ iterator elements(t: TCellSet): PCell {.inline.} =
      inc(i)
    r = r.next

+iterator elementsExcept(t, s: TCellSet): PCell {.inline.} =
+  var r = t.head
+  while r != nil:
+    let ss = CellSetGet(s, r.key)
+    var i = 0
+    while i <= high(r.bits):
+      var w = r.bits[i]
+      if ss != nil:
+        w = w and not ss.bits[i]
+      var j = 0
+      while w != 0:
+        if (w and 1) != 0:
+          yield cast[PCell]((r.key shl PageShift) or
+                              (i shl IntShift +% j) *% MemAlign)
+        inc(j)
+        w = w shr 1
+      inc(i)
+    r = r.next
--- a/lib/system/ecmasys.nim
+++ b/lib/system/ecmasys.nim
@@ -298,8 +298,18 @@ type
    replaceData*: proc (start, len: int, text: cstring) {.nimcall.}
    setAttribute*: proc (name, value: cstring) {.nimcall.}
    setAttributeNode*: proc (attr: ref TNode) {.nimcall.}
+
+when defined(kwin):
+  proc rawEcho {.compilerproc, nostackframe.} =
+    asm """
+      var buf = "";
+      for (var i = 0; i < arguments.length; ++i) {
+        buf += `toEcmaStr`(arguments[i]);
+      }
+      print(buf);
+    """
    
-when defined(nodejs):
+elif defined(nodejs):
  proc ewriteln(x: cstring) = log(x)
  
  proc rawEcho {.compilerproc, nostackframe.} =
--- a/lib/system/gc.nim
+++ b/lib/system/gc.nim
@@ -1,7 +1,7 @@
 #
 #
 #            Nimrod's Runtime Library
-#        (c) Copyright 2012 Andreas Rumpf
+#        (c) Copyright 2013 Andreas Rumpf
 #
 #    See the file "copying.txt", included in this
 #    distribution, for details about the copyright.
@@ -18,6 +18,9 @@
 # for soft real time applications (like games).
 {.push profiler:off.}

+# XXX there is still a slight chance of leaking cycles as we don't add cycle
+# candidates in 'incRef'
+
 const
  CycleIncrease = 2 # is a multiplicative increase
  InitialCycleThreshold = 4*1024*1024 # X MB because cycle checking is slow
@@ -33,19 +36,17 @@ when defined(memProfiler):

 const
  rcIncrement = 0b1000 # so that lowest 3 bits are not touched
-  # NOTE: Most colors are currently unused
  rcBlack = 0b000  # cell is colored black; in use or free
  rcGray = 0b001   # possible member of a cycle
  rcWhite = 0b010  # member of a garbage cycle
  rcPurple = 0b011 # possible root of a cycle
-  rcZct = 0b100    # in ZCT
-  rcRed = 0b101    # Candidate cycle undergoing sigma-computation
-  rcOrange = 0b110 # Candidate cycle awaiting epoch boundary
+  ZctFlag = 0b100  # in ZCT
  rcShift = 3      # shift by rcShift to get the reference counter
-  colorMask = 0b111
+  colorMask = 0b011
 type
  TWalkOp = enum
-    waZctDecRef, waPush, waCycleDecRef
+    waZctDecRef, waPush, waCycleDecRef, waMarkGray, waScan, waScanBlack, 
+    waCollectWhite

  TFinalizer {.compilerproc.} = proc (self: pointer) {.nimcall.}
    # A ref type can have a finalizer that is called before the object's
@@ -88,9 +89,15 @@ template release(gch: TGcHeap) =
  when hasThreadSupport and hasSharedHeap:
    releaseSys(HeapLock)

+template gcAssert(cond: bool, msg: string) =
+  when defined(useGcAssert):
+    if not cond:
+      echo "[GCASSERT] ", msg
+      quit 1
+
 proc addZCT(s: var TCellSeq, c: PCell) {.noinline.} =
-  if (c.refcount and rcZct) == 0:
-    c.refcount = c.refcount and not colorMask or rcZct
+  if (c.refcount and ZctFlag) == 0:
+    c.refcount = c.refcount or ZctFlag
    add(s, c)

 proc cellToUsr(cell: PCell): pointer {.inline.} =
@@ -115,68 +122,22 @@ proc internRefcount(p: pointer): int {.exportc: "getRefcount".} =
 when BitsPerPage mod (sizeof(int)*8) != 0:
  {.error: "(BitsPerPage mod BitsPerUnit) should be zero!".}

-when debugGC:
-  proc writeCell(msg: CString, c: PCell) =
-    var kind = -1
-    if c.typ != nil: kind = ord(c.typ.kind)
-    when leakDetector:
-      c_fprintf(c_stdout, "[GC] %s: %p %d rc=%ld from %s(%ld)\n",
-                msg, c, kind, c.refcount shr rcShift, c.filename, c.line)
-    else:
-      c_fprintf(c_stdout, "[GC] %s: %p %d rc=%ld\n",
-                msg, c, kind, c.refcount shr rcShift)
+template color(c): expr = c.refCount and colorMask
+template setColor(c, col) =
+  when col == rcBlack:
+    c.refcount = c.refCount and not colorMask
+  else:
+    c.refcount = c.refCount and not colorMask or col

-when traceGC:
-  # traceGC is a special switch to enable extensive debugging
-  type
-    TCellState = enum
-      csAllocated, csZctFreed, csCycFreed
-  var
-    states: array[TCellState, TCellSet]
-
-  proc traceCell(c: PCell, state: TCellState) =
-    case state
-    of csAllocated:
-      if c in states[csAllocated]:
-        writeCell("attempt to alloc an already allocated cell", c)
-        sysAssert(false, "traceCell 1")
-      excl(states[csCycFreed], c)
-      excl(states[csZctFreed], c)
-    of csZctFreed:
-      if c in states[csZctFreed]:
-        writeCell("attempt to free zct cell twice", c)
-        sysAssert(false, "traceCell 2")
-      if c in states[csCycFreed]:
-        writeCell("attempt to free with zct, but already freed with cyc", c)
-        sysAssert(false, "traceCell 3")
-      if c notin states[csAllocated]:
-        writeCell("attempt to free not an allocated cell", c)
-        sysAssert(false, "traceCell 4")
-      excl(states[csAllocated], c)
-    of csCycFreed:
-      if c notin states[csAllocated]:
-        writeCell("attempt to free a not allocated cell", c)
-        sysAssert(false, "traceCell 5")
-      if c in states[csCycFreed]:
-        writeCell("attempt to free cyc cell twice", c)
-        sysAssert(false, "traceCell 6")
-      if c in states[csZctFreed]:
-        writeCell("attempt to free with cyc, but already freed with zct", c)
-        sysAssert(false, "traceCell 7")
-      excl(states[csAllocated], c)
-    incl(states[state], c)
-
-  proc writeLeakage() =
-    var z = 0
-    var y = 0
-    var e = 0
-    for c in elements(states[csAllocated]):
-      inc(e)
-      if c in states[csZctFreed]: inc(z)
-      elif c in states[csCycFreed]: inc(y)
-      else: writeCell("leak", c)
-    cfprintf(cstdout, "Allocations: %ld; ZCT freed: %ld; CYC freed: %ld\n",
-             e, z, y)
+proc writeCell(msg: CString, c: PCell) =
+  var kind = -1
+  if c.typ != nil: kind = ord(c.typ.kind)
+  when leakDetector:
+    c_fprintf(c_stdout, "[GC] %s: %p %d rc=%ld from %s(%ld)\n",
+              msg, c, kind, c.refcount shr rcShift, c.filename, c.line)
+  else:
+    c_fprintf(c_stdout, "[GC] %s: %p %d rc=%ld; color=%ld\n",
+              msg, c, kind, c.refcount shr rcShift, c.color)

 template gcTrace(cell, state: expr): stmt {.immediate.} =
  when traceGC: traceCell(cell, state)
@@ -213,7 +174,9 @@ proc rtlAddCycleRoot(c: PCell) {.rtl, inl.} =
  # we MUST access gch as a global here, because this crosses DLL boundaries!
  when hasThreadSupport and hasSharedHeap:
    AcquireSys(HeapLock)
-  incl(gch.cycleRoots, c)
+  if c.color != rcPurple:
+    c.setColor(rcPurple)
+    incl(gch.cycleRoots, c)
  when hasThreadSupport and hasSharedHeap:
    ReleaseSys(HeapLock)

@@ -226,20 +189,22 @@ proc rtlAddZCT(c: PCell) {.rtl, inl.} =
    ReleaseSys(HeapLock)

 proc decRef(c: PCell) {.inline.} =
-  sysAssert(isAllocatedPtr(gch.region, c), "decRef: interiorPtr")
-  sysAssert(c.refcount >=% rcIncrement, "decRef")
+  gcAssert(isAllocatedPtr(gch.region, c), "decRef: interiorPtr")
+  gcAssert(c.refcount >=% rcIncrement, "decRef")
  if --c.refcount:
    rtlAddZCT(c)
  elif canBeCycleRoot(c):
    # unfortunately this is necessary here too, because a cycle might just
    # have been broken up and we could recycle it.
-    rtlAddCycleRoot(c) 
+    rtlAddCycleRoot(c)
+    #writeCell("decRef", c)

 proc incRef(c: PCell) {.inline.} = 
-  sysAssert(isAllocatedPtr(gch.region, c), "incRef: interiorPtr")
-  ++c.refcount
-  if canBeCycleRoot(c):
-    rtlAddCycleRoot(c)
+  gcAssert(isAllocatedPtr(gch.region, c), "incRef: interiorPtr")
+  c.refcount = c.refCount +% rcIncrement and not colorMask
+  #writeCell("incRef", c)
+  #if canBeCycleRoot(c):
+  #  rtlAddCycleRoot(c)

 proc nimGCref(p: pointer) {.compilerProc, inline.} = incRef(usrToCell(p))
 proc nimGCunref(p: pointer) {.compilerProc, inline.} = decRef(usrToCell(p))
@@ -247,7 +212,7 @@ proc nimGCunref(p: pointer) {.compilerProc, inline.} = decRef(usrToCell(p))
 proc nimGCunrefNoCycle(p: pointer) {.compilerProc, inline.} =
  sysAssert(allocInv(gch.region), "begin nimGCunrefNoCycle")
  var c = usrToCell(p)
-  sysAssert(isAllocatedPtr(gch.region, c), "nimGCunrefNoCycle: isAllocatedPtr")
+  gcAssert(isAllocatedPtr(gch.region, c), "nimGCunrefNoCycle: isAllocatedPtr")
  if --c.refcount:
    rtlAddZCT(c)
    sysAssert(allocInv(gch.region), "end nimGCunrefNoCycle 2")
@@ -255,7 +220,7 @@ proc nimGCunrefNoCycle(p: pointer) {.compilerProc, inline.} =

 proc asgnRef(dest: ppointer, src: pointer) {.compilerProc, inline.} =
  # the code generator calls this proc!
-  sysAssert(not isOnStack(dest), "asgnRef")
+  gcAssert(not isOnStack(dest), "asgnRef")
  # BUGFIX: first incRef then decRef!
  if src != nil: incRef(usrToCell(src))
  if dest[] != nil: decRef(usrToCell(dest[]))
@@ -285,14 +250,14 @@ proc unsureAsgnRef(dest: ppointer, src: pointer) {.compilerProc.} =
    if cast[int](dest[]) >=% PageSize: decRef(usrToCell(dest[]))
  else:
    # can't be an interior pointer if it's a stack location!
-    sysAssert(interiorAllocatedPtr(gch.region, dest)==nil, 
-              "stack loc AND interior pointer")
+    gcAssert(interiorAllocatedPtr(gch.region, dest) == nil, 
+             "stack loc AND interior pointer")
  dest[] = src

 proc initGC() =
  when not defined(useNimRtl):
    when traceGC:
-      for i in low(TCellState)..high(TCellState): Init(states[i])
+      for i in low(TCellState)..high(TCellState): init(states[i])
    gch.cycleThreshold = InitialCycleThreshold
    gch.stat.stackScans = 0
    gch.stat.cycleCollections = 0
@@ -303,8 +268,8 @@ proc initGC() =
    # init the rt
    init(gch.zct)
    init(gch.tempStack)
-    Init(gch.cycleRoots)
-    Init(gch.decStack)
+    init(gch.cycleRoots)
+    init(gch.decStack)

 proc forAllSlotsAux(dest: pointer, n: ptr TNimNode, op: TWalkOp) =
  var d = cast[TAddress](dest)
@@ -341,9 +306,9 @@ proc forAllChildrenAux(dest: Pointer, mt: PNimType, op: TWalkOp) =
    else: nil

 proc forAllChildren(cell: PCell, op: TWalkOp) =
-  sysAssert(cell != nil, "forAllChildren: 1")
-  sysAssert(cell.typ != nil, "forAllChildren: 2")
-  sysAssert cell.typ.kind in {tyRef, tySequence, tyString}, "forAllChildren: 3"
+  gcAssert(cell != nil, "forAllChildren: 1")
+  gcAssert(cell.typ != nil, "forAllChildren: 2")
+  gcAssert cell.typ.kind in {tyRef, tySequence, tyString}, "forAllChildren: 3"
  let marker = cell.typ.marker
  if marker != nil:
    marker(cellToUsr(cell), op.int)
@@ -378,7 +343,7 @@ proc addNewObjToZCT(res: PCell, gch: var TGcHeap) {.inline.} =
    template replaceZctEntry(i: expr) =
      c = d[i]
      if c.refcount >=% rcIncrement:
-        c.refcount = c.refcount and not colorMask
+        c.refcount = c.refcount and not ZctFlag
        d[i] = res
        return
    if L > 8:
@@ -399,7 +364,7 @@ proc addNewObjToZCT(res: PCell, gch: var TGcHeap) {.inline.} =
    for i in countdown(L-1, max(0, L-8)):
      var c = d[i]
      if c.refcount >=% rcIncrement:
-        c.refcount = c.refcount and not colorMask
+        c.refcount = c.refcount and not ZctFlag
        d[i] = res
        return
    add(gch.zct, res)
@@ -407,18 +372,19 @@ proc addNewObjToZCT(res: PCell, gch: var TGcHeap) {.inline.} =
 proc rawNewObj(typ: PNimType, size: int, gch: var TGcHeap): pointer =
  # generates a new object and sets its reference counter to 0
  acquire(gch)
-  sysAssert(typ.kind in {tyRef, tyString, tySequence}, "newObj: 1")
+  gcAssert(typ.kind in {tyRef, tyString, tySequence}, "newObj: 1")
  collectCT(gch)
  sysAssert(allocInv(gch.region), "rawNewObj begin")
  var res = cast[PCell](rawAlloc(gch.region, size + sizeof(TCell)))
-  sysAssert((cast[TAddress](res) and (MemAlign-1)) == 0, "newObj: 2")
+  gcAssert((cast[TAddress](res) and (MemAlign-1)) == 0, "newObj: 2")
  # now it is buffered in the ZCT
  res.typ = typ
  when leakDetector and not hasThreadSupport:
    if framePtr != nil and framePtr.prev != nil:
      res.filename = framePtr.prev.filename
      res.line = framePtr.prev.line
-  res.refcount = rcZct # refcount is zero, but mark it to be in the ZCT  
+  # refcount is zero, color is black, but mark it to be in the ZCT
+  res.refcount = ZctFlag
  sysAssert(isAllocatedPtr(gch.region, res), "newObj: 3")
  # its refcount is zero, so add it to the ZCT:
  addNewObjToZCT(res, gch)
@@ -447,7 +413,7 @@ proc newObjRC1(typ: PNimType, size: int): pointer {.compilerRtl.} =
  # generates a new object and sets its reference counter to 1
  sysAssert(allocInv(gch.region), "newObjRC1 begin")
  acquire(gch)
-  sysAssert(typ.kind in {tyRef, tyString, tySequence}, "newObj: 1")
+  gcAssert(typ.kind in {tyRef, tyString, tySequence}, "newObj: 1")
  collectCT(gch)
  sysAssert(allocInv(gch.region), "newObjRC1 after collectCT")
  
@@ -482,7 +448,7 @@ proc growObj(old: pointer, newsize: int, gch: var TGcHeap): pointer =
  collectCT(gch)
  var ol = usrToCell(old)
  sysAssert(ol.typ != nil, "growObj: 1")
-  sysAssert(ol.typ.kind in {tyString, tySequence}, "growObj: 2")
+  gcAssert(ol.typ.kind in {tyString, tySequence}, "growObj: 2")
  sysAssert(allocInv(gch.region), "growObj begin")

  var res = cast[PCell](rawAlloc(gch.region, newsize + sizeof(TCell)))
@@ -499,7 +465,7 @@ proc growObj(old: pointer, newsize: int, gch: var TGcHeap): pointer =
  #  add(gch.zct, res)
  #else: # XXX: what to do here?
  #  decRef(ol)
-  if (ol.refcount and colorMask) == rcZct:
+  if (ol.refcount and ZctFlag) != 0:
    var j = gch.zct.len-1
    var d = gch.zct.d
    while j >= 0: 
@@ -529,74 +495,122 @@ proc growObj(old: pointer, newsize: int): pointer {.rtl.} =

 # ---------------- cycle collector -------------------------------------------

+proc freeCyclicCell(gch: var TGcHeap, c: PCell) =
+  prepareDealloc(c)
+  gcTrace(c, csCycFreed)
+  when logGC: writeCell("cycle collector dealloc cell", c)
+  when reallyDealloc: rawDealloc(gch.region, c)
+  else:
+    gcAssert(c.typ != nil, "freeCyclicCell")
+    zeroMem(c, sizeof(TCell))
+
+proc markGray(s: PCell) =
+  if s.color != rcGray:
+    setColor(s, rcGray)
+    forAllChildren(s, waMarkGray)
+
+proc scanBlack(s: PCell) =
+  s.setColor(rcBlack)
+  forAllChildren(s, waScanBlack)
+
+proc scan(s: PCell) =
+  if s.color == rcGray:
+    if s.refcount >=% rcIncrement:
+      scanBlack(s)
+    else:
+      s.setColor(rcWhite)
+      forAllChildren(s, waScan)
+  
+proc collectWhite(s: PCell) =
+  if s.color == rcWhite and s notin gch.cycleRoots:
+    s.setcolor(rcBlack)
+    forAllChildren(s, waCollectWhite)
+    freeCyclicCell(gch, s)
+
+proc MarkRoots(gch: var TGcHeap) =
+  var tabSize = 0
+  for s in elements(gch.cycleRoots):
+    #writeCell("markRoot", s)
+    inc tabSize
+    if s.color == rcPurple and s.refCount >=% rcIncrement:
+      markGray(s)
+    else:
+      excl(gch.cycleRoots, s)
+      # (s.color == rcBlack and rc == 0) as 1 condition:
+      if s.refcount == 0:
+        freeCyclicCell(gch, s)
+  gch.stat.cycleTableSize = max(gch.stat.cycleTableSize, tabSize)
+
 proc doOperation(p: pointer, op: TWalkOp) =
  if p == nil: return
  var c: PCell = usrToCell(p)
-  sysAssert(c != nil, "doOperation: 1")
-  case op # faster than function pointers because of easy prediction
+  gcAssert(c != nil, "doOperation: 1")
+  # the 'case' should be faster than function pointers because of easy
+  # prediction:
+  case op
  of waZctDecRef:
    #if not isAllocatedPtr(gch.region, c):
    #  return
    #  c_fprintf(c_stdout, "[GC] decref bug: %p", c) 
-    sysAssert(isAllocatedPtr(gch.region, c), "decRef: waZctDecRef")
-    sysAssert(c.refcount >=% rcIncrement, "doOperation 2")
-    c.refcount = c.refcount -% rcIncrement
+    gcAssert(isAllocatedPtr(gch.region, c), "decRef: waZctDecRef")
+    gcAssert(c.refcount >=% rcIncrement, "doOperation 2")
+    #c.refcount = c.refcount -% rcIncrement
    when logGC: writeCell("decref (from doOperation)", c)
-    if c.refcount <% rcIncrement: addZCT(gch.zct, c)
+    decRef(c)
+    #if c.refcount <% rcIncrement: addZCT(gch.zct, c)
  of waPush:
    add(gch.tempStack, c)
  of waCycleDecRef:
-    sysAssert(c.refcount >=% rcIncrement, "doOperation 3")
+    gcAssert(c.refcount >=% rcIncrement, "doOperation 3")
    c.refcount = c.refcount -% rcIncrement
+  of waMarkGray:
+    gcAssert(c.refcount >=% rcIncrement, "waMarkGray")
+    c.refcount = c.refcount -% rcIncrement
+    markGray(c)
+  of waScan: scan(c)
+  of waScanBlack:
+    c.refcount = c.refcount +% rcIncrement
+    if c.color != rcBlack:
+      scanBlack(c)
+  of waCollectWhite: collectWhite(c)

 proc nimGCvisit(d: pointer, op: int) {.compilerRtl.} =
  doOperation(d, TWalkOp(op))

-# we now use a much simpler and non-recursive algorithm for cycle removal
-proc collectCycles(gch: var TGcHeap) =
-  var tabSize = 0
-  for c in elements(gch.cycleRoots):
-    inc(tabSize)
-    forallChildren(c, waCycleDecRef)
-  if tabSize == 0: return
-  gch.stat.cycleTableSize = max(gch.stat.cycleTableSize, tabSize)
+proc CollectZCT(gch: var TGcHeap): bool
+
+proc collectRoots(gch: var TGcHeap) =
+  for s in elements(gch.cycleRoots):
+    excl(gch.cycleRoots, s)
+    collectWhite(s)
+
+proc collectCycles(gch: var TGcHeap) =
+  # ensure the ZCT 'color' is not used:
+  while gch.zct.len > 0: discard collectZCT(gch)
+  markRoots(gch)
+  # scanRoots:
+  for s in elements(gch.cycleRoots): scan(s)
+  collectRoots(gch)

-  # restore reference counts (a depth-first traversal is needed):
-  var marker: TCellSet
-  Init(marker)
-  for c in elements(gch.cycleRoots):
-    if c.refcount >=% rcIncrement:
-      if not containsOrIncl(marker, c):
-        gch.tempStack.len = 0
-        forAllChildren(c, waPush)
-        while gch.tempStack.len > 0:
-          dec(gch.tempStack.len)
-          var d = gch.tempStack.d[gch.tempStack.len]
-          d.refcount = d.refcount +% rcIncrement
-          if d in gch.cycleRoots and not containsOrIncl(marker, d):
-            forAllChildren(d, waPush)
-  Deinit(marker)
-  # remove cycles:
-  for c in elements(gch.cycleRoots):
-    if c.refcount <% rcIncrement:
-      gch.tempStack.len = 0
-      forAllChildren(c, waPush)
-      while gch.tempStack.len > 0:
-        dec(gch.tempStack.len)
-        var d = gch.tempStack.d[gch.tempStack.len]
-        if d.refcount <% rcIncrement:
-          if d notin gch.cycleRoots: # d is leaf of c and not part of cycle
-            addZCT(gch.zct, d)
-            when logGC: writeCell("add to ZCT (from cycle collector)", d)
-      prepareDealloc(c)
-      gcTrace(c, csCycFreed)
-      when logGC: writeCell("cycle collector dealloc cell", c)
-      when reallyDealloc: rawDealloc(gch.region, c)
-      else:
-        sysAssert(c.typ != nil, "collectCycles")
-        zeroMem(c, sizeof(TCell))
  Deinit(gch.cycleRoots)
  Init(gch.cycleRoots)
+  # alive cycles need to be kept in 'cycleRoots' if they are referenced
+  # from the stack; otherwise the write barrier will add the cycle root again
+  # anyway:
+  when false:
+    var d = gch.decStack.d
+    var cycleRootsLen = 0
+    for i in 0..gch.decStack.len-1:
+      var c = d[i]
+      gcAssert isAllocatedPtr(gch.region, c), "addBackStackRoots"
+      gcAssert c.refcount >=% rcIncrement, "addBackStackRoots: dead cell"
+      if canBeCycleRoot(c):
+        #if c notin gch.cycleRoots: 
+        inc cycleRootsLen
+        incl(gch.cycleRoots, c)
+      gcAssert c.typ != nil, "addBackStackRoots 2"
+    if cycleRootsLen != 0:
+      cfprintf(cstdout, "cycle roots: %ld\n", cycleRootsLen)

 proc gcMark(gch: var TGcHeap, p: pointer) {.inline.} =
  # the addresses are not as cells on the stack, so turn them to cells:
@@ -778,9 +792,9 @@ proc CollectZCT(gch: var TGcHeap): bool =
    var c = gch.zct.d[0]
    sysAssert(isAllocatedPtr(gch.region, c), "CollectZCT: isAllocatedPtr")
    # remove from ZCT:
-    sysAssert((c.refcount and rcZct) == rcZct, "collectZCT")
+    gcAssert((c.refcount and ZctFlag) == ZctFlag, "collectZCT")
    
-    c.refcount = c.refcount and not colorMask
+    c.refcount = c.refcount and not ZctFlag
    gch.zct.d[0] = gch.zct.d[L[] - 1]
    dec(L[])
    when withRealtime: dec steps
@@ -809,22 +823,22 @@ proc CollectZCT(gch: var TGcHeap): bool =
        if gch.maxPause > 0:
          let duration = getticks() - t0
          # the GC's measuring is not accurate and needs some cleanup actions 
-          # (stack unmarking), so subtract some short amount of time in to
+          # (stack unmarking), so subtract some short amount of time in
          # order to miss deadlines less often:
          if duration >= gch.maxPause - 50_000:
            return false
  result = true

-proc unmarkStackAndRegisters(gch: var TGcHeap) = 
+proc unmarkStackAndRegisters(gch: var TGcHeap) =
  var d = gch.decStack.d
  for i in 0..gch.decStack.len-1:
    sysAssert isAllocatedPtr(gch.region, d[i]), "unmarkStackAndRegisters"
-    # decRef(d[i]) inlined: cannot create a cycle and must not acquire lock
-    var c = d[i]
+    decRef(d[i])
+    #var c = d[i]
    # XXX no need for an atomic dec here:
-    if --c.refcount:
-      addZCT(gch.zct, c)
-    sysAssert c.typ != nil, "unmarkStackAndRegisters 2"
+    #if --c.refcount:
+    #  addZCT(gch.zct, c)
+    #sysAssert c.typ != nil, "unmarkStackAndRegisters 2"
  gch.decStack.len = 0

 proc collectCTBody(gch: var TGcHeap) =
@@ -843,7 +857,7 @@ proc collectCTBody(gch: var TGcHeap) =
    when cycleGC:
      if getOccupiedMem(gch.region) >= gch.cycleThreshold or alwaysCycleGC:
        collectCycles(gch)
-        discard collectZCT(gch)
+        #discard collectZCT(gch)
        inc(gch.stat.cycleCollections)
        gch.cycleThreshold = max(InitialCycleThreshold, getOccupiedMem() *
                                 cycleIncrease)
@@ -929,7 +943,6 @@ when not defined(useNimRtl):
             "[GC] max cycle table size: " & $gch.stat.cycleTableSize & "\n" &
             "[GC] max stack size: " & $gch.stat.maxStackSize & "\n" &
             "[GC] max pause time [ms]: " & $(gch.stat.maxPause div 1000_000)
-    when traceGC: writeLeakage()
    GC_enable()

 {.pop.}
--- a/lib/system/gc_ms.nim
+++ b/lib/system/gc_ms.nim
@@ -0,0 +1,504 @@
+#
+#
+#            Nimrod's Runtime Library
+#        (c) Copyright 2013 Andreas Rumpf
+#
+#    See the file "copying.txt", included in this
+#    distribution, for details about the copyright.
+#
+
+# A simple mark&sweep garbage collector for Nimrod.
+{.push profiler:off.}
+
+const
+  InitialThreshold = 4*1024*1024 # X MB because marking&sweeping is slow
+
+template mulThreshold(x): expr {.immediate.} = x * 2
+
+when defined(memProfiler):
+  proc nimProfile(requestedSize: int)
+
+type
+  TWalkOp = enum
+    waMarkGlobal,  # we need to mark conservatively for global marker procs
+                   # as these may refer to a global var and not to a thread
+                   # local 
+    waMarkPrecise  # fast precise marking
+
+  TFinalizer {.compilerproc.} = proc (self: pointer) {.nimcall.}
+    # A ref type can have a finalizer that is called before the object's
+    # storage is freed.
+  
+  TGlobalMarkerProc = proc () {.nimcall.}
+
+  TGcStat = object
+    collections: int         # number of performed full collections
+    maxThreshold: int        # max threshold that has been set
+    maxStackSize: int        # max stack size
+    freedObjects: int        # max entries in cycle table  
+  
+  TGcHeap = object           # this contains the zero count and
+                             # non-zero count table
+    stackBottom: pointer
+    cycleThreshold: int
+    allocated, marked: TCellSet
+    tempStack: TCellSeq      # temporary stack for recursion elimination
+    recGcLock: int           # prevent recursion via finalizers; no thread lock
+    region: TMemRegion       # garbage collected region
+    stat: TGcStat
+
+var
+  gch {.rtlThreadVar.}: TGcHeap
+
+when not defined(useNimRtl):
+  InstantiateForRegion(gch.region)
+
+template acquire(gch: TGcHeap) = 
+  when hasThreadSupport and hasSharedHeap:
+    AcquireSys(HeapLock)
+
+template release(gch: TGcHeap) = 
+  when hasThreadSupport and hasSharedHeap:
+    releaseSys(HeapLock)
+
+template gcAssert(cond: bool, msg: string) =
+  when defined(useGcAssert):
+    if not cond:
+      echo "[GCASSERT] ", msg
+      quit 1
+
+proc cellToUsr(cell: PCell): pointer {.inline.} =
+  # convert object (=pointer to refcount) to pointer to userdata
+  result = cast[pointer](cast[TAddress](cell)+%TAddress(sizeof(TCell)))
+
+proc usrToCell(usr: pointer): PCell {.inline.} =
+  # convert pointer to userdata to object (=pointer to refcount)
+  result = cast[PCell](cast[TAddress](usr)-%TAddress(sizeof(TCell)))
+
+proc canbeCycleRoot(c: PCell): bool {.inline.} =
+  result = ntfAcyclic notin c.typ.flags
+
+proc extGetCellType(c: pointer): PNimType {.compilerproc.} =
+  # used for code generation concerning debugging
+  result = usrToCell(c).typ
+
+proc unsureAsgnRef(dest: ppointer, src: pointer) {.inline.} =
+  dest[] = src
+
+proc internRefcount(p: pointer): int {.exportc: "getRefcount".} =
+  result = 0
+
+var
+  globalMarkersLen: int
+  globalMarkers: array[0.. 7_000, TGlobalMarkerProc]
+
+proc nimRegisterGlobalMarker(markerProc: TGlobalMarkerProc) {.compilerProc.} =
+  if globalMarkersLen <= high(globalMarkers):
+    globalMarkers[globalMarkersLen] = markerProc
+    inc globalMarkersLen
+  else:
+    echo "[GC] cannot register global variable; too many global variables"
+    quit 1
+
+# this that has to equals zero, otherwise we have to round up UnitsPerPage:
+when BitsPerPage mod (sizeof(int)*8) != 0:
+  {.error: "(BitsPerPage mod BitsPerUnit) should be zero!".}
+
+# forward declarations:
+proc collectCT(gch: var TGcHeap)
+proc IsOnStack*(p: pointer): bool {.noinline.}
+proc forAllChildren(cell: PCell, op: TWalkOp)
+proc doOperation(p: pointer, op: TWalkOp)
+proc forAllChildrenAux(dest: Pointer, mt: PNimType, op: TWalkOp)
+# we need the prototype here for debugging purposes
+
+proc prepareDealloc(cell: PCell) =
+  if cell.typ.finalizer != nil:
+    # the finalizer could invoke something that
+    # allocates memory; this could trigger a garbage
+    # collection. Since we are already collecting we
+    # prevend recursive entering here by a lock.
+    # XXX: we should set the cell's children to nil!
+    inc(gch.recGcLock)
+    (cast[TFinalizer](cell.typ.finalizer))(cellToUsr(cell))
+    dec(gch.recGcLock)
+
+proc nimGCref(p: pointer) {.compilerProc, inline.} = 
+  # we keep it from being collected by pretending it's not even allocated:
+  excl(gch.allocated, usrToCell(p))
+proc nimGCunref(p: pointer) {.compilerProc, inline.} = 
+  incl(gch.allocated, usrToCell(p))
+
+proc initGC() =
+  when not defined(useNimRtl):
+    gch.cycleThreshold = InitialThreshold
+    gch.stat.collections = 0
+    gch.stat.maxThreshold = 0
+    gch.stat.maxStackSize = 0
+    init(gch.tempStack)
+    Init(gch.allocated)
+    init(gch.marked)
+
+proc forAllSlotsAux(dest: pointer, n: ptr TNimNode, op: TWalkOp) =
+  var d = cast[TAddress](dest)
+  case n.kind
+  of nkSlot: forAllChildrenAux(cast[pointer](d +% n.offset), n.typ, op)
+  of nkList:
+    for i in 0..n.len-1:
+      forAllSlotsAux(dest, n.sons[i], op)
+  of nkCase:
+    var m = selectBranch(dest, n)
+    if m != nil: forAllSlotsAux(dest, m, op)
+  of nkNone: sysAssert(false, "forAllSlotsAux")
+
+proc forAllChildrenAux(dest: Pointer, mt: PNimType, op: TWalkOp) =
+  var d = cast[TAddress](dest)
+  if dest == nil: return # nothing to do
+  if ntfNoRefs notin mt.flags:
+    case mt.Kind
+    of tyRef, tyString, tySequence: # leaf:
+      doOperation(cast[ppointer](d)[], op)
+    of tyObject, tyTuple:
+      forAllSlotsAux(dest, mt.node, op)
+    of tyArray, tyArrayConstr, tyOpenArray:
+      for i in 0..(mt.size div mt.base.size)-1:
+        forAllChildrenAux(cast[pointer](d +% i *% mt.base.size), mt.base, op)
+    else: nil
+
+proc forAllChildren(cell: PCell, op: TWalkOp) =
+  gcAssert(cell != nil, "forAllChildren: 1")
+  gcAssert(cell.typ != nil, "forAllChildren: 2")
+  gcAssert cell.typ.kind in {tyRef, tySequence, tyString}, "forAllChildren: 3"
+  let marker = cell.typ.marker
+  if marker != nil:
+    marker(cellToUsr(cell), op.int)
+  else:
+    case cell.typ.Kind
+    of tyRef: # common case
+      forAllChildrenAux(cellToUsr(cell), cell.typ.base, op)
+    of tySequence:
+      var d = cast[TAddress](cellToUsr(cell))
+      var s = cast[PGenericSeq](d)
+      if s != nil:
+        for i in 0..s.len-1:
+          forAllChildrenAux(cast[pointer](d +% i *% cell.typ.base.size +%
+            GenericSeqSize), cell.typ.base, op)
+    else: nil
+
+proc rawNewObj(typ: PNimType, size: int, gch: var TGcHeap): pointer =
+  # generates a new object and sets its reference counter to 0
+  acquire(gch)
+  gcAssert(typ.kind in {tyRef, tyString, tySequence}, "newObj: 1")
+  collectCT(gch)
+  var res = cast[PCell](rawAlloc(gch.region, size + sizeof(TCell)))
+  gcAssert((cast[TAddress](res) and (MemAlign-1)) == 0, "newObj: 2")
+  # now it is buffered in the ZCT
+  res.typ = typ
+  when leakDetector and not hasThreadSupport:
+    if framePtr != nil and framePtr.prev != nil:
+      res.filename = framePtr.prev.filename
+      res.line = framePtr.prev.line
+  res.refcount = 0
+  release(gch)
+  incl(gch.allocated, res)
+  result = cellToUsr(res)
+
+{.pop.}
+
+proc newObj(typ: PNimType, size: int): pointer {.compilerRtl.} =
+  result = rawNewObj(typ, size, gch)
+  zeroMem(result, size)
+  when defined(memProfiler): nimProfile(size)
+
+proc newSeq(typ: PNimType, len: int): pointer {.compilerRtl.} =
+  # `newObj` already uses locks, so no need for them here.
+  let size = addInt(mulInt(len, typ.base.size), GenericSeqSize)
+  result = newObj(typ, size)
+  cast[PGenericSeq](result).len = len
+  cast[PGenericSeq](result).reserved = len
+  when defined(memProfiler): nimProfile(size)
+
+proc newObjRC1(typ: PNimType, size: int): pointer {.compilerRtl.} =
+  result = rawNewObj(typ, size, gch)
+  zeroMem(result, size)
+  when defined(memProfiler): nimProfile(size)
+  
+proc newSeqRC1(typ: PNimType, len: int): pointer {.compilerRtl.} =
+  let size = addInt(mulInt(len, typ.base.size), GenericSeqSize)
+  result = newObj(typ, size)
+  cast[PGenericSeq](result).len = len
+  cast[PGenericSeq](result).reserved = len
+  when defined(memProfiler): nimProfile(size)
+  
+proc growObj(old: pointer, newsize: int, gch: var TGcHeap): pointer =
+  acquire(gch)
+  collectCT(gch)
+  var ol = usrToCell(old)
+  sysAssert(ol.typ != nil, "growObj: 1")
+  gcAssert(ol.typ.kind in {tyString, tySequence}, "growObj: 2")
+  
+  var res = cast[PCell](rawAlloc(gch.region, newsize + sizeof(TCell)))
+  var elemSize = 1
+  if ol.typ.kind != tyString: elemSize = ol.typ.base.size
+  
+  var oldsize = cast[PGenericSeq](old).len*elemSize + GenericSeqSize
+  copyMem(res, ol, oldsize + sizeof(TCell))
+  zeroMem(cast[pointer](cast[TAddress](res)+% oldsize +% sizeof(TCell)),
+          newsize-oldsize)
+  sysAssert((cast[TAddress](res) and (MemAlign-1)) == 0, "growObj: 3")
+  excl(gch.allocated, ol)
+  when reallyDealloc: rawDealloc(gch.region, ol)
+  else:
+    zeroMem(ol, sizeof(TCell))
+  incl(gch.allocated, res)
+  release(gch)
+  result = cellToUsr(res)
+  when defined(memProfiler): nimProfile(newsize-oldsize)
+
+proc growObj(old: pointer, newsize: int): pointer {.rtl.} =
+  result = growObj(old, newsize, gch)
+
+{.push profiler:off.}
+
+# ----------------- collector -----------------------------------------------
+
+proc mark(gch: var TGcHeap, c: PCell) =
+  incl(gch.marked, c)
+  gcAssert gch.tempStack.len == 0, "stack not empty!"
+  forAllChildren(c, waMarkPrecise)
+  while gch.tempStack.len > 0:
+    dec gch.tempStack.len
+    var d = gch.tempStack.d[gch.tempStack.len]
+    if not containsOrIncl(gch.marked, d):
+      forAllChildren(d, waMarkPrecise)
+
+proc doOperation(p: pointer, op: TWalkOp) =
+  if p == nil: return
+  var c: PCell = usrToCell(p)
+  gcAssert(c != nil, "doOperation: 1")
+  case op
+  of waMarkGlobal:
+    when hasThreadSupport:
+      # could point to a cell which we don't own and don't want to touch/trace
+      if isAllocatedPtr(gch.region, c):
+        mark(gch, c)
+    else:
+      mark(gch, c)
+  of waMarkPrecise: add(gch.tempStack, c)
+
+proc nimGCvisit(d: pointer, op: int) {.compilerRtl.} =
+  doOperation(d, TWalkOp(op))
+
+proc freeCyclicCell(gch: var TGcHeap, c: PCell) =
+  inc gch.stat.freedObjects
+  prepareDealloc(c)
+  when reallyDealloc: rawDealloc(gch.region, c)
+  else:
+    gcAssert(c.typ != nil, "freeCyclicCell")
+    zeroMem(c, sizeof(TCell))
+
+proc sweep(gch: var TGcHeap) =
+  for c in gch.allocated.elementsExcept(gch.marked):
+    gch.allocated.excl(c)
+    freeCyclicCell(gch, c)
+
+proc markGlobals(gch: var TGcHeap) =
+  for i in 0 .. < globalMarkersLen: globalMarkers[i]()
+
+proc gcMark(gch: var TGcHeap, p: pointer) {.inline.} =
+  # the addresses are not as cells on the stack, so turn them to cells:
+  var cell = usrToCell(p)
+  var c = cast[TAddress](cell)
+  if c >% PageSize:
+    # fast check: does it look like a cell?
+    var objStart = cast[PCell](interiorAllocatedPtr(gch.region, cell))
+    if objStart != nil:
+      mark(gch, objStart)
+  
+# ----------------- stack management --------------------------------------
+#  inspired from Smart Eiffel
+
+when defined(sparc):
+  const stackIncreases = false
+elif defined(hppa) or defined(hp9000) or defined(hp9000s300) or
+     defined(hp9000s700) or defined(hp9000s800) or defined(hp9000s820):
+  const stackIncreases = true
+else:
+  const stackIncreases = false
+
+when not defined(useNimRtl):
+  {.push stack_trace: off.}
+  proc setStackBottom(theStackBottom: pointer) =
+    #c_fprintf(c_stdout, "stack bottom: %p;\n", theStackBottom)
+    # the first init must be the one that defines the stack bottom:
+    if gch.stackBottom == nil: gch.stackBottom = theStackBottom
+    else:
+      var a = cast[TAddress](theStackBottom) # and not PageMask - PageSize*2
+      var b = cast[TAddress](gch.stackBottom)
+      #c_fprintf(c_stdout, "old: %p new: %p;\n",gch.stackBottom,theStackBottom)
+      when stackIncreases:
+        gch.stackBottom = cast[pointer](min(a, b))
+      else:
+        gch.stackBottom = cast[pointer](max(a, b))
+  {.pop.}
+
+proc stackSize(): int {.noinline.} =
+  var stackTop {.volatile.}: pointer
+  result = abs(cast[int](addr(stackTop)) - cast[int](gch.stackBottom))
+
+when defined(sparc): # For SPARC architecture.
+  proc isOnStack(p: pointer): bool =
+    var stackTop {.volatile.}: pointer
+    stackTop = addr(stackTop)
+    var b = cast[TAddress](gch.stackBottom)
+    var a = cast[TAddress](stackTop)
+    var x = cast[TAddress](p)
+    result = a <=% x and x <=% b
+
+  proc markStackAndRegisters(gch: var TGcHeap) {.noinline, cdecl.} =
+    when defined(sparcv9):
+      asm  """"flushw \n" """
+    else:
+      asm  """"ta      0x3   ! ST_FLUSH_WINDOWS\n" """
+
+    var
+      max = gch.stackBottom
+      sp: PPointer
+      stackTop: array[0..1, pointer]
+    sp = addr(stackTop[0])
+    # Addresses decrease as the stack grows.
+    while sp <= max:
+      gcMark(gch, sp[])
+      sp = cast[ppointer](cast[TAddress](sp) +% sizeof(pointer))
+
+elif defined(ELATE):
+  {.error: "stack marking code is to be written for this architecture".}
+
+elif stackIncreases:
+  # ---------------------------------------------------------------------------
+  # Generic code for architectures where addresses increase as the stack grows.
+  # ---------------------------------------------------------------------------
+  proc isOnStack(p: pointer): bool =
+    var stackTop {.volatile.}: pointer
+    stackTop = addr(stackTop)
+    var a = cast[TAddress](gch.stackBottom)
+    var b = cast[TAddress](stackTop)
+    var x = cast[TAddress](p)
+    result = a <=% x and x <=% b
+
+  var
+    jmpbufSize {.importc: "sizeof(jmp_buf)", nodecl.}: int
+      # a little hack to get the size of a TJmpBuf in the generated C code
+      # in a platform independant way
+
+  proc markStackAndRegisters(gch: var TGcHeap) {.noinline, cdecl.} =
+    var registers: C_JmpBuf
+    if c_setjmp(registers) == 0'i32: # To fill the C stack with registers.
+      var max = cast[TAddress](gch.stackBottom)
+      var sp = cast[TAddress](addr(registers)) +% jmpbufSize -% sizeof(pointer)
+      # sp will traverse the JMP_BUF as well (jmp_buf size is added,
+      # otherwise sp would be below the registers structure).
+      while sp >=% max:
+        gcMark(gch, cast[ppointer](sp)[])
+        sp = sp -% sizeof(pointer)
+
+else:
+  # ---------------------------------------------------------------------------
+  # Generic code for architectures where addresses decrease as the stack grows.
+  # ---------------------------------------------------------------------------
+  proc isOnStack(p: pointer): bool =
+    var stackTop {.volatile.}: pointer
+    stackTop = addr(stackTop)
+    var b = cast[TAddress](gch.stackBottom)
+    var a = cast[TAddress](stackTop)
+    var x = cast[TAddress](p)
+    result = a <=% x and x <=% b
+
+  proc markStackAndRegisters(gch: var TGcHeap) {.noinline, cdecl.} =
+    # We use a jmp_buf buffer that is in the C stack.
+    # Used to traverse the stack and registers assuming
+    # that 'setjmp' will save registers in the C stack.
+    type PStackSlice = ptr array [0..7, pointer]
+    var registers: C_JmpBuf
+    if c_setjmp(registers) == 0'i32: # To fill the C stack with registers.
+      var max = cast[TAddress](gch.stackBottom)
+      var sp = cast[TAddress](addr(registers))
+      # loop unrolled:
+      while sp <% max - 8*sizeof(pointer):
+        gcMark(gch, cast[PStackSlice](sp)[0])
+        gcMark(gch, cast[PStackSlice](sp)[1])
+        gcMark(gch, cast[PStackSlice](sp)[2])
+        gcMark(gch, cast[PStackSlice](sp)[3])
+        gcMark(gch, cast[PStackSlice](sp)[4])
+        gcMark(gch, cast[PStackSlice](sp)[5])
+        gcMark(gch, cast[PStackSlice](sp)[6])
+        gcMark(gch, cast[PStackSlice](sp)[7])
+        sp = sp +% sizeof(pointer)*8
+      # last few entries:
+      while sp <=% max:
+        gcMark(gch, cast[ppointer](sp)[])
+        sp = sp +% sizeof(pointer)
+
+# ----------------------------------------------------------------------------
+# end of non-portable code
+# ----------------------------------------------------------------------------
+
+proc collectCTBody(gch: var TGcHeap) =
+  gch.stat.maxStackSize = max(gch.stat.maxStackSize, stackSize())
+  prepareForInteriorPointerChecking(gch.region)
+  markStackAndRegisters(gch)
+  markGlobals(gch)
+  sweep(gch)
+  
+  inc(gch.stat.collections)
+  deinit(gch.marked)
+  init(gch.marked)
+  gch.cycleThreshold = max(InitialThreshold, getOccupiedMem().mulThreshold)
+  gch.stat.maxThreshold = max(gch.stat.maxThreshold, gch.cycleThreshold)
+  sysAssert(allocInv(gch.region), "collectCT: end")
+  
+proc collectCT(gch: var TGcHeap) =
+  if getOccupiedMem(gch.region) >= gch.cycleThreshold and gch.recGcLock == 0:
+    collectCTBody(gch)
+
+when not defined(useNimRtl):
+  proc GC_disable() = 
+    when hasThreadSupport and hasSharedHeap:
+      atomicInc(gch.recGcLock, 1)
+    else:
+      inc(gch.recGcLock)
+  proc GC_enable() =
+    if gch.recGcLock > 0: 
+      when hasThreadSupport and hasSharedHeap:
+        atomicDec(gch.recGcLock, 1)
+      else:
+        dec(gch.recGcLock)
+
+  proc GC_setStrategy(strategy: TGC_Strategy) = nil
+
+  proc GC_enableMarkAndSweep() =
+    gch.cycleThreshold = InitialThreshold
+
+  proc GC_disableMarkAndSweep() =
+    gch.cycleThreshold = high(gch.cycleThreshold)-1
+    # set to the max value to suppress the cycle detector
+
+  proc GC_fullCollect() =
+    acquire(gch)
+    var oldThreshold = gch.cycleThreshold
+    gch.cycleThreshold = 0 # forces cycle collection
+    collectCT(gch)
+    gch.cycleThreshold = oldThreshold
+    release(gch)
+
+  proc GC_getStatistics(): string =
+    GC_disable()
+    result = "[GC] total memory: " & $getTotalMem() & "\n" &
+             "[GC] occupied memory: " & $getOccupiedMem() & "\n" &
+             "[GC] collections: " & $gch.stat.collections & "\n" &
+             "[GC] max threshold: " & $gch.stat.maxThreshold & "\n" &
+             "[GC] freed objects: " & $gch.stat.freedObjects & "\n" &
+             "[GC] max stack size: " & $gch.stat.maxStackSize & "\n"
+    GC_enable()
+
+{.pop.}
--- a/lib/system/mmdisp.nim
+++ b/lib/system/mmdisp.nim
@@ -309,6 +309,9 @@ else:
    sysAssert(sizeof(TCell) == sizeof(TFreeCell), "sizeof TFreeCell")
  when compileOption("gc", "v2"):
    include "system/gc2"
+  elif defined(gcMarkAndSweep):
+    # XXX use 'compileOption' here
+    include "system/gc_ms"
  else:
    include "system/gc"
  
--- a/lib/system/sysstr.nim
+++ b/lib/system/sysstr.nim
@@ -202,7 +202,8 @@ proc setLengthSeq(seq: PGenericSeq, elemSize, newLen: int): PGenericSeq {.
                               GenericSeqSize))
  elif newLen < result.len:
    # we need to decref here, otherwise the GC leaks!
-    when not defined(boehmGC) and not defined(nogc):
+    when not defined(boehmGC) and not defined(nogc) and 
+         not defined(gcMarkAndSweep):
      when compileOption("gc", "v2"):
        for i in newLen..result.len-1:
          let len0 = gch.tempStack.len
--- a/tests/gc/cycleleak.nim
+++ b/tests/gc/cycleleak.nim
@@ -5,38 +5,52 @@ discard """
 type
  Module = object
    nodes*: seq[PNode]
+    id: int

  PModule = ref Module

  Node = object
    owner*: PModule
    data*: array[0..200, char] # some fat to drain memory faster
+    id: int

  PNode = ref Node

+var
+  gid: int
+
+when false:
+  proc finalizeNode(x: PNode) =
+    echo "node id: ", x.id
+  proc finalizeModule(x: PModule) =
+    echo "module id: ", x.id
+
 proc newNode(owner: PModule): PNode =
  new(result)
  result.owner = owner
+  inc gid
+  result.id = gid

 proc compileModule: PModule =
  new(result)
  result.nodes = @[]
  for i in 0..100:
    result.nodes.add newNode(result)
+  inc gid
+  result.id = gid

 var gModuleCache: PModule

 proc loop =
-  for i in 0..10000:
+  for i in 0..1000:
    gModuleCache = compileModule()
    gModuleCache = nil
    GC_fullCollect()

-  if getOccupiedMem() > 300_000:
-    echo "still a leak! ", getOccupiedMem()
-    quit(1)
-  else:
-    echo "no leak: ", getOccupiedMem()
+    if getOccupiedMem() > 9_000_000:
+      echo "still a leak! ", getOccupiedMem()
+      quit(1)
+  echo "no leak: ", getOccupiedMem()

 loop()

--- a/tests/gc/gcleak.nim
+++ b/tests/gc/gcleak.nim
@@ -12,7 +12,9 @@ type
 proc MakeObj(): TTestObj =
  result.x = "Hello"

-for i in 1 .. 100_000_000:
+for i in 1 .. 1_000_000:
+  when defined(gcMarkAndSweep):
+    GC_fullcollect()
  var obj = MakeObj()
  if getOccupiedMem() > 300_000: quit("still a leak!")
 #  echo GC_getstatistics()
--- a/tests/gc/gcleak2.nim
+++ b/tests/gc/gcleak2.nim
@@ -15,7 +15,9 @@ proc MakeObj(): TTestObj =
  result.s = @[1,2,3]

 proc inProc() = 
-  for i in 1 .. 100_000_000:
+  for i in 1 .. 1_000_000:
+    when defined(gcMarkAndSweep):
+      GC_fullcollect()
    var obj: TTestObj
    obj = MakeObj()
    if getOccupiedMem() > 300_000: quit("still a leak!")
--- a/tests/gc/weakrefs.nim
+++ b/tests/gc/weakrefs.nim
@@ -40,6 +40,8 @@ proc main =
  for i in 0 .. s.high:
    s[i] = register(create())
  # test that we have at least 80% unreachable weak objects by now:
+  when defined(gcMarkAndSweep):
+    GC_fullcollect()
  var unreachable = 0
  for i in 0 .. s.high:
    if access(s[i]) == nil: inc unreachable
--- a/tests/reject/trecinca.nim
+++ b/tests/reject/trecinca.nim
@@ -1,5 +1,5 @@
 discard """
-  file: "trecincb.nim"
+  file: "tests/reject/trecincb.nim"
  line: 9
  errormsg: "recursive dependency: 'trecincb.nim'"
 """
--- a/tests/reject/trecincb.nim
+++ b/tests/reject/trecincb.nim
@@ -1,7 +1,7 @@
 discard """
  file: "trecincb.nim"
  line: 9
-  errormsg: "recursive dependency: 'trecincb.nim'"
+  errormsg: "recursive dependency: 'tests/reject/trecincb.nim'"
 """
 # Test recursive includes

--- a/tests/specials.nim
+++ b/tests/specials.nim
@@ -119,6 +119,10 @@ proc runGcTests(r: var TResults, options: string) =
    runSingleTest(r, "tests/gc" / filename, options & " -d:release")
    runSingleTest(r, "tests/gc" / filename, options &
                  " -d:release -d:useRealtimeGC")
+    runSingleTest(r, "tests/gc" / filename, options &
+                  " --gc:markAndSweep")
+    runSingleTest(r, "tests/gc" / filename, options &
+                  " -d:release --gc:markAndSweep")
  
  test "gcbench"
  test "gcleak"
--- a/todo.txt
+++ b/todo.txt
@@ -4,6 +4,8 @@ version 0.9.2
 - implement constructors + full 'not nil' checking
 - ``restrict`` pragma + backend support
 - fix: 'result' is not properly cleaned for NRVO
+- document NimMain and check whether it works for threading
+

 version 0.9.4
 =============
--- a/web/news.txt
+++ b/web/news.txt
@@ -10,6 +10,12 @@ Version 0.9.2 has been released! Get it `here <download.html>`_.
 Bugfixes
 --------

+- The old GC never collected cycles correctly. Fixed but it can cause
+  performance regressions. However you can deactivate the cycle collector 
+  with ``GC_disableMarkAndSweep`` and run it explicitly at an appropriate time 
+  or not at all. There is also a new GC you can activate 
+  with ``--gc:markAndSweep`` which does not have this problem but is slower in
+  general.


 Library Additions
@@ -18,6 +24,8 @@ Library Additions
 - Added ``system.onRaise`` to support a condition system.
 - Added ``macros.quote`` for AST quasi-quoting.
 - Added ``system.unsafeNew`` to support hacky variable length objects.
+- There is a new experimental mark&sweep GC which can be faster (or much 
+  slower) than the default GC. Enable with ``--gc:markAndSweep``.


 Changes affecting backwards compatibility