Merge pull request #631 from mflamer/master

LockFree Hash Table 0.1

4 files changed, 850 insertions, 50 deletions

diff --git a/lib/pure/collections/LockFreeHash.nim b/lib/pure/collections/LockFreeHash.nim
new file mode 100644
index 000000000..d3a91763a
--- /dev/null
+++ b/lib/pure/collections/LockFreeHash.nim
@@ -0,0 +1,581 @@
+#nimrod c -t:-march=i686 --cpu:amd64 --threads:on -d:release lockfreehash.nim
+
+import baseutils, unsigned, math, hashes
+
+
+
+const
+  minTableSize = 8
+  reProbeLimit = 12
+  minCopyWork = 4096
+  intSize = sizeof(int)
+  
+
+
+when sizeof(int) == 4: # 32bit
+  type 
+    TRaw = range[0..1073741823]
+    ## The range of uint values that can be stored directly in a value slot
+    ## when on a 32 bit platform
+
+elif sizeof(int) == 8: # 64bit
+  type
+    TRaw = range[0..4611686018427387903]
+    ## The range of uint values that can be stored directly in a value slot
+    ## when on a 64 bit platform
+else: echo("unsupported platform")
+  
+type  
+  TEntry = tuple
+    key: int
+    value: int
+
+  TEntryArr = ptr array[0..10_000_000, TEntry]
+  
+  PConcTable[K,V] = ptr object {.pure.}
+    len: int
+    used: int
+    active: int
+    copyIdx: int
+    copyDone: int
+    next: PConcTable[K,V]    
+    data: TEntryArr
+
+
+proc setVal[K,V](table: var PConcTable[K,V], key: int, val: int,
+  expVal: int, match: bool): int 
+
+#------------------------------------------------------------------------------
+
+# Create a new table
+proc newLFTable*[K,V](size: int = minTableSize): PConcTable[K,V]  =
+  let 
+    dataLen = max(nextPowerOfTwo(size), minTableSize)   
+    dataSize = dataLen*sizeof(TEntry) 
+    dataMem = allocShared0(dataSize) 
+    tableSize = 7 * intSize
+    tableMem = allocShared0(tableSize)
+    table = cast[PConcTable[K,V]](tableMem) 
+  table.len = dataLen
+  table.used = 0
+  table.active = 0
+  table.copyIdx = 0
+  table.copyDone = 0
+  table.next = nil
+  table.data = cast[TEntryArr](dataMem)
+  result = table
+
+#------------------------------------------------------------------------------  
+
+# Delete a table
+proc deleteConcTable[K,V](tbl: PConcTable[K,V]) =
+  deallocShared(tbl.data) 
+  deallocShared(tbl)
+
+#------------------------------------------------------------------------------  
+
+proc `[]`[K,V](table: var PConcTable[K,V], i: int): var TEntry {.inline.} =
+  table.data[i]
+
+#------------------------------------------------------------------------------
+# State flags stored in ptr
+
+
+proc pack[T](x: T): int {.inline.} =
+  result = (cast[int](x) shl 2)
+  #echo("packKey ",cast[int](x) , " -> ", result)
+
+# Pop the flags off returning a 4 byte aligned ptr to our Key or Val 
+proc pop(x: int): int  {.inline.} =
+  result = x and 0xFFFFFFFC'i32
+
+# Pop the raw value off of our Key or Val 
+proc popRaw(x: int): int  {.inline.} =
+  result = x shr 2  
+
+# Pop the flags off returning a 4 byte aligned ptr to our Key or Val 
+proc popPtr[V](x: int): ptr V  {.inline.} =
+  result = cast[ptr V](pop(x))
+  #echo("popPtr " & $x & " -> " & $cast[int](result))
+
+# Ghost (sentinel)
+# K or V is no longer valid use new table
+const Ghost = 0xFFFFFFFC
+proc isGhost(x: int): bool {.inline.} =
+  result = x == 0xFFFFFFFC  
+
+# Tombstone 
+# applied to V = K is dead 
+proc isTomb(x: int): bool {.inline.} = 
+  result = (x and 0x00000002) != 0
+
+proc setTomb(x: int): int {.inline.} =
+  result = x or 0x00000002
+
+# Prime
+# K or V is in new table copied from old  
+proc isPrime(x: int): bool {.inline.} = 
+  result = (x and 0x00000001) != 0
+
+proc setPrime(x: int): int {.inline.} =
+  result = x or 0x00000001
+
+#------------------------------------------------------------------------------  
+
+##This is for i32 only need to override for i64
+proc hashInt(x: int):int {.inline.} = 
+  var h = uint32(x) #shr 2'u32  
+  h = h xor (h shr 16'u32)
+  h *= 0x85ebca6b'u32
+  h = h xor (h shr 13'u32)
+  h *= 0xc2b2ae35'u32
+  h = h xor (h shr 16'u32)  
+  result = int(h)
+
+#------------------------------------------------------------------------------
+
+proc resize[K,V](self: PConcTable[K,V]): PConcTable[K,V] =
+  var next = atomic_load_n(self.next.addr, ATOMIC_RELAXED)
+  #echo("next = " & $cast[int](next))
+  if next != nil:
+    #echo("A new table already exists, copy in progress")
+    return next
+  var
+    oldLen = atomic_load_n(self.len.addr, ATOMIC_RELAXED)     
+    newTable = newLFTable[K,V](oldLen*2)
+    success = atomic_compare_exchange_n(self.next.addr, next.addr, newTable,
+                  false, ATOMIC_RELAXED, ATOMIC_RELAXED)
+  if not success:
+    echo("someone beat us to it! delete table we just created and return his " & $cast[int](next))
+    deleteConcTable(newTable) 
+    return next     
+  else:
+    echo("Created New Table! " & $cast[int](newTable) & " Size = " & $newTable.len)
+    return newTable
+  
+
+#------------------------------------------------------------------------------
+#proc keyEQ[K](key1: ptr K, key2: ptr K): bool {.inline.} = 
+proc keyEQ[K](key1: int, key2: int): bool {.inline.} = 
+  result = false
+  when K is TRaw:
+    if key1 == key2: 
+      result = true
+  else:
+    var 
+      p1 = popPtr[K](key1)
+      p2 = popPtr[K](key2)
+    if p1 != nil and p2 != nil:  
+      if cast[int](p1) == cast[int](p2):
+        return true
+      if p1[] == p2[]:
+        return true
+
+#------------------------------------------------------------------------------
+
+#proc tableFull(self: var PConcTable[K,V]) : bool {.inline.} =
+
+
+#------------------------------------------------------------------------------
+
+proc copySlot[K,V](idx: int, oldTbl: var PConcTable[K,V], newTbl: var PConcTable[K,V]): bool =
+  #echo("Copy idx " & $idx)
+  var 
+    oldVal = 0
+    oldkey = 0   
+    ok = false
+  result = false
+  #Block the key so no other threads waste time here
+  while not ok:
+    ok = atomic_compare_exchange_n(oldTbl[idx].key.addr, oldKey.addr, 
+      setTomb(oldKey), false, ATOMIC_RELAXED, ATOMIC_RELAXED)
+  #echo("oldKey was = " & $oldKey & "  set it to tomb " & $setTomb(oldKey)) 
+  #Prevent new values from appearing in the old table by priming 
+  oldVal = atomic_load_n(oldTbl[idx].value.addr, ATOMIC_RELAXED)
+  while not isPrime(oldVal):
+    var box = if oldVal == NULL or isTomb(oldVal) : oldVal.setTomb.setPrime 
+      else: oldVal.setPrime 
+    if atomic_compare_exchange_n(oldTbl[idx].value.addr, oldVal.addr, 
+      box, false, ATOMIC_RELAXED, ATOMIC_RELAXED):
+      if isPrime(box) and isTomb(box):     
+        return true
+      oldVal = box
+      break
+  #echo("oldVal was = ", oldVal, "  set it to prime ", box)
+  if isPrime(oldVal) and isTomb(oldVal): 
+    #when not (K is TRaw):
+    #  deallocShared(popPtr[K](oldKey)) 
+    return false
+  if isTomb(oldVal): 
+    echo("oldVal is Tomb!!!, should not happen")
+  if pop(oldVal) != NULL:  
+    result = setVal(newTbl, pop(oldKey), pop(oldVal), NULL, true) == NULL
+  if result: 
+    #echo("Copied a Slot! idx= " & $idx & " key= " & $oldKey & " val= " & $oldVal)    
+  else: 
+    #echo("copy slot failed")    
+  # Our copy is done so we disable the old slot
+  while not ok:
+    ok = atomic_compare_exchange_n(oldTbl[idx].value.addr, oldVal.addr, 
+      oldVal.setTomb.setPrime , false, ATOMIC_RELAXED, ATOMIC_RELAXED)
+  #echo("disabled old slot") 
+  #echo"---------------------" 
+
+#------------------------------------------------------------------------------
+
+proc promote[K,V](table: var PConcTable[K,V]) =
+  var
+    newData = atomic_load_n(table.next.data.addr, ATOMIC_RELAXED)
+    newLen = atomic_load_n(table.next.len.addr, ATOMIC_RELAXED) 
+    newUsed = atomic_load_n(table.next.used.addr, ATOMIC_RELAXED)
+
+  deallocShared(table.data)
+  atomic_store_n(table.data.addr, newData, ATOMIC_RELAXED)
+  atomic_store_n(table.len.addr, newLen, ATOMIC_RELAXED)
+  atomic_store_n(table.used.addr, newUsed, ATOMIC_RELAXED)
+  atomic_store_n(table.copyIdx.addr, 0, ATOMIC_RELAXED)
+  atomic_store_n(table.copyDone.addr, 0, ATOMIC_RELAXED)
+  deallocShared(table.next)
+  atomic_store_n(table.next.addr, nil, ATOMIC_RELAXED) 
+  echo("new table swapped!")
+
+#------------------------------------------------------------------------------
+ 
+proc checkAndPromote[K,V](table: var PConcTable[K,V], workDone: int): bool =  
+  var 
+    oldLen = atomic_load_n(table.len.addr, ATOMIC_RELAXED)
+    copyDone = atomic_load_n(table.copyDone.addr, ATOMIC_RELAXED)
+    ok: bool
+  result = false  
+  if workDone > 0:
+    #echo("len to copy =" & $oldLen)
+    #echo("copyDone + workDone = " & $copyDone & " + " & $workDone) 
+    while not ok:
+      ok = atomic_compare_exchange_n(table.copyDone.addr, copyDone.addr, 
+        copyDone + workDone, false, ATOMIC_RELAXED, ATOMIC_RELAXED)
+    #if ok: echo("set copyDone")                 
+    # If the copy is done we can promote this table 
+    if copyDone + workDone >= oldLen:
+      # Swap new data
+      #echo("work is done!")        
+      table.promote
+      result = true
+    
+#------------------------------------------------------------------------------
+
+proc copySlotAndCheck[K,V](table: var PConcTable[K,V], idx: int):
+  PConcTable[K,V] =
+  var
+    newTable = cast[PConcTable[K,V]](atomic_load_n(table.next.addr, ATOMIC_RELAXED))
+  result = newTable  
+  if newTable != nil and copySlot(idx, table, newTable): 
+    #echo("copied a single slot, idx = " & $idx)   
+    if checkAndPromote(table, 1): return table
+  
+
+#------------------------------------------------------------------------------
+  
+proc helpCopy[K,V](table: var PConcTable[K,V]): PConcTable[K,V] =
+  var
+    newTable = cast[PConcTable[K,V]](atomic_load_n(table.next.addr, ATOMIC_RELAXED)) 
+  result = newTable   
+  if newTable != nil:     
+    var 
+      oldLen = atomic_load_n(table.len.addr, ATOMIC_RELAXED) 
+      copyDone = atomic_load_n(table.copyDone.addr, ATOMIC_RELAXED)
+      copyIdx = 0
+      work = min(oldLen, minCopyWork)
+      #panicStart = -1
+      workDone = 0
+    if copyDone < oldLen:
+      var ok: bool
+      while not ok:
+        ok = atomic_compare_exchange_n(table.copyIdx.addr, copyIdx.addr, 
+          copyIdx + work, false, ATOMIC_RELAXED, ATOMIC_RELAXED)
+      #echo("copy idx = ", copyIdx)  
+      for i in 0..work-1:
+        var idx = (copyIdx + i) and (oldLen - 1)        
+        if copySlot(idx, table, newTable):
+          workDone += 1
+      if workDone > 0:
+        #echo("did work ", workDone, " on thread ", cast[int](myThreadID[pointer]()))
+        if checkAndPromote(table, workDone): return table
+    # In case a thread finished all the work then got stalled before promotion 
+    if checkAndPromote(table, 0): return table
+    
+  
+
+#------------------------------------------------------------------------------
+
+proc setVal[K,V](table: var PConcTable[K,V], key: int, val: int,
+  expVal: int, match: bool): int =
+  #echo("-try set- in table ", " key = ", (popPtr[K](key)[]), " val = ", val)   
+  when K is TRaw: 
+    var idx = hashInt(key)    
+  else:
+    var idx = popPtr[K](key)[].hash                
+  var    
+    nextTable: PConcTable[K,V]   
+    probes = 1
+  # spin until we find a key slot or build and jump to next table    
+  while true:     
+    idx = idx and (table.len - 1) 
+    #echo("try set idx = " & $idx & "for" & $key)
+    var
+      probedKey = NULL     
+      openKey = atomic_compare_exchange_n(table[idx].key.addr, probedKey.addr, 
+        key, false, ATOMIC_RELAXED, ATOMIC_RELAXED) 
+    if openKey:
+      if val.isTomb:
+        #echo("val was tomb, bail, no reason to set an open slot to tomb")
+        return val
+      #increment used slots 
+      #echo("found an open slot, total used = " & 
+      #$atomic_add_fetch(table.used.addr, 1, ATOMIC_RELAXED))
+      discard atomic_add_fetch(table.used.addr, 1, ATOMIC_RELAXED)
+      break # We found an open slot 
+    #echo("set idx ", idx, " key = ", key, " probed = ", probedKey)       
+    if keyEQ[K](probedKey, key):
+      #echo("we found the matching slot")    
+      break # We found a matching slot      
+    if (not(expVal != NULL and match)) and (probes >= reProbeLimit or key.isTomb):
+      if key.isTomb: echo("Key is Tombstone")
+      #if probes >= reProbeLimit: echo("Too much probing " & $probes)
+      #echo("try to resize")
+      #create next bigger table
+      nextTable = resize(table)
+      #help do some copying
+      #echo("help copy old table to new")       
+      nextTable = helpCopy(table)      
+      #now setVal in the new table instead
+      #echo("jumping to next table to set val")       
+      return setVal(nextTable, key, val, expVal, match)      
+    else:
+      idx += 1
+      probes += 1
+  # Done spinning for a new slot
+  var oldVal = atomic_load_n(table[idx].value.addr, ATOMIC_RELAXED)     
+  if val == oldVal:
+    #echo("this val is alredy in the slot") 
+    return oldVal
+  nextTable = atomic_load_n(table.next.addr, ATOMIC_SEQ_CST)  
+  if nextTable == nil and 
+    ((oldVal == NULL and 
+    (probes >= reProbeLimit or table.used / table.len > 0.8)) or
+    (isPrime(oldVal))):
+    if table.used / table.len > 0.8: echo("resize because usage ratio = " &
+      $(table.used / table.len))
+    if isPrime(oldVal): echo("old val isPrime, should be a rare mem ordering event")
+    nextTable = resize(table)
+  if nextTable != nil:
+    #echo("tomb old slot then set in new table") 
+    nextTable = copySlotAndCheck(table,idx)
+    return setVal(nextTable, key, val, expVal, match)
+  # Finaly ready to add new val to table
+  while true:
+    if match and oldVal != expVal:
+      #echo("set failed, no match  oldVal= " & $oldVal & " expVal= " & $expVal)
+      return oldVal
+    if atomic_compare_exchange_n(table[idx].value.addr, oldVal.addr, 
+        val, false, ATOMIC_RELEASE, ATOMIC_RELAXED):
+      #echo("val set at table " & $cast[int](table))
+      if expVal != NULL:
+        if (oldVal == NULL or isTomb(oldVal)) and not isTomb(val):
+          discard atomic_add_fetch(table.active.addr, 1, ATOMIC_RELAXED)
+        elif not (oldVal == NULL or isTomb(oldVal)) and isTomb(val):
+          discard atomic_add_fetch(table.active.addr, -1, ATOMIC_RELAXED)
+      if oldVal == NULL and expVal != NULL:
+        return setTomb(oldVal)
+      else: return oldVal
+    if isPrime(oldVal):
+      nextTable = copySlotAndCheck(table, idx)
+      return setVal(nextTable, key, val, expVal, match)
+
+#------------------------------------------------------------------------------
+
+proc getVal[K,V](table: var PConcTable[K,V], key: int): int = 
+  #echo("-try get-  key = " & $key)
+  when K is TRaw: 
+    var idx = hashInt(key)
+  else:
+    var idx = popPtr[K](key)[].hash 
+    #echo("get idx ", idx)    
+  var    
+    probes = 0
+    val: int  
+  while true:
+    idx = idx and (table.len - 1)        
+    var 
+      newTable: PConcTable[K,V] # = atomic_load_n(table.next.addr, ATOMIC_ACQUIRE)
+      probedKey = atomic_load_n(table[idx].key.addr, ATOMIC_SEQ_CST)    
+    if keyEQ[K](probedKey, key):
+      #echo("found key after ", probes+1)
+      val = atomic_load_n(table[idx].value.addr, ATOMIC_ACQUIRE)
+      if not isPrime(val):
+        if isTomb(val):
+          #echo("val was tomb but not prime") 
+          return NULL
+        else:
+          #echo("-GotIt- idx = ", idx, " key = ", key, " val ", val ) 
+          return val
+      else:
+        newTable = copySlotAndCheck(table, idx)
+        return getVal(newTable, key) 
+    else:
+      #echo("probe ", probes, " idx = ", idx, " key = ", key, " found ", probedKey )    
+      if probes >= reProbeLimit*4 or key.isTomb:
+        if newTable == nil:
+          #echo("too many probes and no new table ", key, "  ", idx )
+          return NULL
+        else: 
+          newTable = helpCopy(table)
+          return getVal(newTable, key)
+      idx += 1
+      probes += 1
+
+#------------------------------------------------------------------------------
+   
+#proc set*(table: var PConcTable[TRaw,TRaw], key: TRaw, val: TRaw) =
+#  discard setVal(table, pack(key), pack(key), NULL, false)
+
+#proc set*[V](table: var PConcTable[TRaw,V], key: TRaw, val: ptr V) =
+#  discard setVal(table, pack(key), cast[int](val), NULL, false)
+
+proc set*[K,V](table: var PConcTable[K,V], key: var K, val: var V) =
+  when not (K is TRaw): 
+    var newKey = cast[int](copyShared(key))
+  else: 
+    var newKey = pack(key)
+  when not (V is TRaw): 
+    var newVal = cast[int](copyShared(val))
+  else: 
+    var newVal = pack(val)
+  var oldPtr = pop(setVal(table, newKey, newVal, NULL, false))
+    #echo("oldPtr = ", cast[int](oldPtr), " newPtr = ", cast[int](newPtr))
+  when not (V is TRaw): 
+    if newVal != oldPtr and oldPtr != NULL: 
+      deallocShared(cast[ptr V](oldPtr))
+  
+ 
+
+proc get*[K,V](table: var PConcTable[K,V], key: var K): V =
+  when not (V is TRaw):
+    when not (K is TRaw):
+      return popPtr[V](getVal(table, cast[int](key.addr)))[]
+    else: 
+      return popPtr[V](getVal(table, pack(key)))[]
+  else:
+    when not (K is TRaw):
+      return popRaw(getVal(table, cast[int](key.addr)))
+    else: 
+      return popRaw(getVal(table, pack(key)))   
+
+
+
+
+
+
+
+
+
+    
+
+#proc `[]`[K,V](table: var PConcTable[K,V], key: K): PEntry[K,V] {.inline.} =
+#  getVal(table, key)
+
+#proc `[]=`[K,V](table: var PConcTable[K,V], key: K, val: V): PEntry[K,V] {.inline.} =
+#  setVal(table, key, val)
+
+
+
+
+
+
+#Tests ----------------------------
+when isMainModule:
+  import locks, times, mersenne
+  
+  const 
+    numTests  = 100000
+    numThreads = 10
+
+
+    
+  type
+    TTestObj = tuple
+      thr: int      
+      f0: int
+      f1: int
+
+    TData = tuple[k: string,v: TTestObj]
+    PDataArr = array[0..numTests-1, TData]
+    Dict = PConcTable[string,TTestObj]
+    
+  var 
+    thr: array[0..numThreads-1, TThread[Dict]]
+    
+    table = newLFTable[string,TTestObj](8)        
+    rand = newMersenneTwister(2525)
+
+  proc createSampleData(len: int): PDataArr = 
+    #result = cast[PDataArr](allocShared0(sizeof(TData)*numTests))             
+    for i in 0..len-1:
+      result[i].k = "mark" & $(i+1)
+      #echo("mark" & $(i+1), " ", hash("mark" & $(i+1)))      
+      result[i].v.thr = 0
+      result[i].v.f0 = i+1 
+      result[i].v.f1 = 0       
+      #echo("key = " & $(i+1) & " Val ptr = " & $cast[int](result[i].v.addr))
+
+
+
+  proc threadProc(tp: Dict) {.thread.} = 
+    var t = cpuTime();     
+    for i in 1..numTests:
+      var key = "mark" & $(i)
+      var got = table.get(key)                         
+      got.thr = cast[int](myThreadID[pointer]())
+      got.f1 = got.f1 + 1                
+      table.set(key, got)
+    t = cpuTime() - t
+    echo t             
+  
+
+  var testData = createSampleData(numTests)
+
+  for i in 0..numTests-1:
+    table.set(testData[i].k, testData[i].v)
+    
+  var i = 0
+  while i < numThreads:
+    createThread(thr[i], threadProc, table)
+    i += 1
+
+  joinThreads(thr) 
+
+
+
+    
+
+  var fails = 0
+
+  for i in 0..numTests-1: 
+    var got = table.get(testData[i].k) 
+    if got.f0 != i+1 or got.f1 != numThreads:
+      fails += 1
+      echo(got)
+
+  echo("Failed read or write = ", fails)
+     
+
+  #for i in 1..numTests:
+  #  echo(i, " = ", hashInt(i) and 8191)
+
+  deleteConcTable(table)
+
+
+
+
+
+
+
diff --git a/lib/pure/collections/baseutils.nim b/lib/pure/collections/baseutils.nim
new file mode 100644
index 000000000..565a89ccb
--- /dev/null
+++ b/lib/pure/collections/baseutils.nim
@@ -0,0 +1,41 @@
+
+
+
+#------------------------------------------------------------------------------
+## Useful Constants
+const NULL* = 0
+
+
+#------------------------------------------------------------------------------
+## Memory Utility Functions
+
+proc newHeap*[T](): ptr T =
+  result = cast[ptr T](alloc0(sizeof(T))) 
+
+proc copyNew*[T](x: var T): ptr T =
+  var 
+    size = sizeof(T)    
+    mem = alloc(size)  
+  copyMem(mem, x.addr, size)  
+  return cast[ptr T](mem)
+
+proc copyTo*[T](val: var T, dest: int) = 
+  copyMem(pointer(dest), val.addr, sizeof(T))    
+
+proc allocType*[T](): pointer = alloc(sizeof(T)) 
+
+proc newShared*[T](): ptr T =
+  result = cast[ptr T](allocShared0(sizeof(T))) 
+
+proc copyShared*[T](x: var T): ptr T =
+  var 
+    size = sizeof(T)    
+    mem = allocShared(size)  
+  copyMem(mem, x.addr, size)  
+  return cast[ptr T](mem)
+
+#------------------------------------------------------------------------------
+## Pointer arithmetic 
+
+proc `+`*(p: pointer, i: int): pointer {.inline.} =
+  cast[pointer](cast[int](p) + i)
\ No newline at end of file
diff --git a/lib/pure/mersenne.nim b/lib/pure/mersenne.nim
new file mode 100644
index 000000000..2b12cce73
--- /dev/null
+++ b/lib/pure/mersenne.nim
@@ -0,0 +1,39 @@
+import unsigned
+
+type
+  TMersenneTwister* = object
+    mt: array[0..623, uint32]
+    index: int
+
+proc newMersenneTwister*(seed: int): TMersenneTwister =   
+  result.index = 0
+  result.mt[0]= uint32(seed)
+  for i in 1..623'u32:
+    result.mt[i]= (0x6c078965'u32 * (result.mt[i-1] xor (result.mt[i-1] shr 30'u32)) + i)
+
+proc generateNumbers(m: var TMersenneTwister) =
+  for i in 0..623:
+    var y = (m.mt[i] and 0x80000000'u32) + (m.mt[(i+1) mod 624] and 0x7fffffff'u32)
+    m.mt[i] = m.mt[(i+397) mod 624] xor uint32(y shr 1'u32)
+    if (y mod 2'u32) != 0:
+     m.mt[i] = m.mt[i] xor 0x9908b0df'u32
+
+proc getNum*(m: var TMersenneTwister): int =
+  if m.index == 0:
+    generateNumbers(m)
+  var y = m.mt[m.index]
+  y = y xor (y shr 11'u32)
+  y = y xor ((7'u32 shl y) and 0x9d2c5680'u32)
+  y = y xor ((15'u32 shl y) and 0xefc60000'u32)
+  y = y xor (y shr 18'u32)
+  m.index = (m.index+1) mod 624
+  return int(y)
+
+
+
+# Test
+when isMainModule:
+  var mt = newMersenneTwister(2525)
+
+  for i in 0..99:
+    echo mt.getNum
\ No newline at end of file
diff --git a/lib/system/atomics.nim b/lib/system/atomics.nim
index 623f8d0d2..36185e0a8 100644
--- a/lib/system/atomics.nim
+++ b/lib/system/atomics.nim
@@ -9,68 +9,207 @@
 
 # Atomic operations for Nimrod.
 
-when (defined(gcc) or defined(llvm_gcc)) and hasThreadSupport and 
-    not defined(windows):
-  proc sync_add_and_fetch(p: var int, val: int): int {.
-    importc: "__sync_add_and_fetch", nodecl.}
-  proc sync_sub_and_fetch(p: var int, val: int): int {.
-    importc: "__sync_sub_and_fetch", nodecl.}
+when (defined(gcc) or defined(llvm_gcc)) and hasThreadSupport: 
+  
+  type 
+    AtomMemModel* = enum
+      ATOMIC_RELAXED, 
+    ## No barriers or synchronization. 
+      ATOMIC_CONSUME, 
+    ## Data dependency only for both barrier and synchronization with another thread.
+      ATOMIC_ACQUIRE, 
+    ## Barrier to hoisting of code and synchronizes with release (or stronger) 
+    ## semantic stores from another thread.
+      ATOMIC_RELEASE,
+    ## Barrier to sinking of code and synchronizes with acquire (or stronger) 
+    ## semantic loads from another thread. 
+      ATOMIC_ACQ_REL,
+    ## Full barrier in both directions and synchronizes with acquire loads 
+    ## and release stores in another thread.
+      ATOMIC_SEQ_CST
+    ## Full barrier in both directions and synchronizes with acquire loads 
+    ## and release stores in all threads.
+
+    TAtomType* = TNumber|pointer|ptr|char
+    ## Type Class representing valid types for use with atomic procs
+
+  proc atomic_load_n*[T: TAtomType](p: ptr T, mem: AtomMemModel): T {.
+    importc: "__atomic_load_n", nodecl.}
+    ## This proc implements an atomic load operation. It returns the contents at p.
+    ## ATOMIC_RELAXED, ATOMIC_SEQ_CST, ATOMIC_ACQUIRE, ATOMIC_CONSUME.
+
+  proc atomic_load*[T: TAtomType](p: ptr T, ret: ptr T, mem: AtomMemModel) {.
+    importc: "__atomic_load", nodecl.}  
+    ## This is the generic version of an atomic load. It returns the contents at p in ret.
+
+  proc atomic_store_n*[T: TAtomType](p: ptr T, val: T, mem: AtomMemModel) {.
+    importc: "__atomic_store_n", nodecl.} 
+    ## This proc implements an atomic store operation. It writes val at p.
+    ## ATOMIC_RELAXED, ATOMIC_SEQ_CST, and ATOMIC_RELEASE.
+
+  proc atomic_store*[T: TAtomType](p: ptr T, val: ptr T, mem: AtomMemModel) {.
+    importc: "__atomic_store", nodecl.}
+    ## This is the generic version of an atomic store. It stores the value of val at p
+
+  proc atomic_exchange_n*[T: TAtomType](p: ptr T, val: T, mem: AtomMemModel): T {.
+    importc: "__atomic_exchange_n", nodecl.}
+    ## This proc implements an atomic exchange operation. It writes val at p, 
+    ## and returns the previous contents at p.
+    ## ATOMIC_RELAXED, ATOMIC_SEQ_CST, ATOMIC_ACQUIRE, ATOMIC_RELEASE, ATOMIC_ACQ_REL
+
+  proc atomic_exchange*[T: TAtomType](p: ptr T, val: ptr T, ret: ptr T, mem: AtomMemModel) {.
+    importc: "__atomic_exchange", nodecl.}
+    ## This is the generic version of an atomic exchange. It stores the contents at val at p. 
+    ## The original value at p is copied into ret.
+
+  proc atomic_compare_exchange_n*[T: TAtomType](p: ptr T, expected: ptr T, desired: T,
+    weak: bool, success_memmodel: AtomMemModel, failure_memmodel: AtomMemModel): bool {.
+    importc: "__atomic_compare_exchange_n ", nodecl.} 
+    ## This proc implements an atomic compare and exchange operation. This compares the
+    ## contents at p with the contents at expected and if equal, writes desired at p. 
+    ## If they are not equal, the current contents at p is written into expected. 
+    ## Weak is true for weak compare_exchange, and false for the strong variation. 
+    ## Many targets only offer the strong variation and ignore the parameter. 
+    ## When in doubt, use the strong variation.
+    ## True is returned if desired is written at p and the execution is considered 
+    ## to conform to the memory model specified by success_memmodel. There are no 
+    ## restrictions on what memory model can be used here. False is returned otherwise, 
+    ## and the execution is considered to conform to failure_memmodel. This memory model 
+    ## cannot be __ATOMIC_RELEASE nor __ATOMIC_ACQ_REL. It also cannot be a stronger model 
+    ## than that specified by success_memmodel.
+
+  proc atomic_compare_exchange*[T: TAtomType](p: ptr T, expected: ptr T, desired: ptr T,
+    weak: bool, success_memmodel: AtomMemModel, failure_memmodel: AtomMemModel): bool {.
+    importc: "__atomic_compare_exchange_n ", nodecl.}  
+    ## This proc implements the generic version of atomic_compare_exchange. 
+    ## The proc is virtually identical to atomic_compare_exchange_n, except the desired 
+    ## value is also a pointer. 
+
+  ## Perform the operation return the new value, all memory models are valid 
+  proc atomic_add_fetch*[T: TAtomType](p: ptr T, val: T, mem: AtomMemModel): T {.
+    importc: "__atomic_add_fetch", nodecl.}
+  proc atomic_sub_fetch*[T: TAtomType](p: ptr T, val: T, mem: AtomMemModel): T {.
+    importc: "__atomic_sub_fetch", nodecl.}
+  proc atomic_or_fetch*[T: TAtomType](p: ptr T, val: T, mem: AtomMemModel): T {.
+    importc: "__atomic_or_fetch ", nodecl.}
+  proc atomic_and_fetch*[T: TAtomType](p: ptr T, val: T, mem: AtomMemModel): T {.
+    importc: "__atomic_and_fetch", nodecl.}
+  proc atomic_xor_fetch*[T: TAtomType](p: ptr T, val: T, mem: AtomMemModel): T {.   
+    importc: "__atomic_xor_fetch", nodecl.}
+  proc atomic_nand_fetch*[T: TAtomType](p: ptr T, val: T, mem: AtomMemModel): T {.   
+    importc: "__atomic_nand_fetch ", nodecl.} 
+
+  ## Perform the operation return the old value, all memory models are valid 
+  proc atomic_fetch_add*[T: TAtomType](p: ptr T, val: T, mem: AtomMemModel): T {.
+    importc: "__atomic_fetch_add ", nodecl.}
+  proc atomic_fetch_sub*[T: TAtomType](p: ptr T, val: T, mem: AtomMemModel): T {.  
+    importc: "__atomic_fetch_sub ", nodecl.}
+  proc atomic_fetch_or*[T: TAtomType](p: ptr T, val: T, mem: AtomMemModel): T {.  
+    importc: "__atomic_fetch_or ", nodecl.}
+  proc atomic_fetch_and*[T: TAtomType](p: ptr T, val: T, mem: AtomMemModel): T {.   
+    importc: "__atomic_fetch_and ", nodecl.}
+  proc atomic_fetch_xor*[T: TAtomType](p: ptr T, val: T, mem: AtomMemModel): T {.  
+    importc: "__atomic_fetch_xor ", nodecl.}
+  proc atomic_fetch_nand*[T: TAtomType](p: ptr T, val: T, mem: AtomMemModel): T {.  
+    importc: "__atomic_fetch_nand", nodecl.} 
+
+  proc atomic_test_and_set*(p: pointer, mem: AtomMemModel): bool {.  
+    importc: "__atomic_test_and_set ", nodecl.} 
+    ## This built-in function performs an atomic test-and-set operation on the byte at p. 
+    ## The byte is set to some implementation defined nonzero “set” value and the return
+    ## value is true if and only if the previous contents were “set”.
+    ## All memory models are valid.
+
+  proc atomic_clear*(p: pointer, mem: AtomMemModel) {.  
+    importc: "__atomic_clear", nodecl.}
+    ## This built-in function performs an atomic clear operation at p. 
+    ## After the operation, at p contains 0.
+    ## ATOMIC_RELAXED, ATOMIC_SEQ_CST, ATOMIC_RELEASE
+
+  proc atomic_thread_fence*(mem: AtomMemModel) {.  
+    importc: "__atomic_thread_fence ", nodecl.}
+    ## This built-in function acts as a synchronization fence between threads based 
+    ## on the specified memory model. All memory orders are valid.
+
+  proc atomic_signal_fence*(mem: AtomMemModel) {.  
+    importc: "__atomic_signal_fence  ", nodecl.}
+    ## This built-in function acts as a synchronization fence between a thread and 
+    ## signal handlers based in the same thread. All memory orders are valid.
+
+  proc atomic_always_lock_free*(size: int, p: pointer): bool {.  
+    importc: "__atomic_always_lock_free   ", nodecl.}
+    ## This built-in function returns true if objects of size bytes always generate 
+    ## lock free atomic instructions for the target architecture. size must resolve 
+    ## to a compile-time constant and the result also resolves to a compile-time constant.
+    ## ptr is an optional pointer to the object that may be used to determine alignment. 
+    ## A value of 0 indicates typical alignment should be used. The compiler may also 
+    ## ignore this parameter.
+
+  proc atomic_is_lock_free*(size: int, p: pointer): bool {.  
+    importc: "__atomic_is_lock_free    ", nodecl.}
+    ## This built-in function returns true if objects of size bytes always generate 
+    ## lock free atomic instructions for the target architecture. If it is not known 
+    ## to be lock free a call is made to a runtime routine named __atomic_is_lock_free.
+    ## ptr is an optional pointer to the object that may be used to determine alignment. 
+    ## A value of 0 indicates typical alignment should be used. The compiler may also 
+    ## ignore this parameter.
+
+
+
 elif defined(vcc) and hasThreadSupport:
-  proc sync_add_and_fetch(p: var int, val: int): int {.
+  proc add_and_fetch*(p: ptr int, val: int): int {.
     importc: "NimXadd", nodecl.}
 else:
-  proc sync_add_and_fetch(p: var int, val: int): int {.inline.} =
-    inc(p, val)
-    result = p
+  proc add_and_fetch*(p: ptr int, val: int): int {.inline.} =
+    inc(p[], val)
+    result = p[]
+
+
+
+
+# atomic compare and swap (CAS) funcitons to implement lock-free algorithms  
+      
+#if defined(windows) and not defined(gcc) and hasThreadSupport:
+#    proc InterlockedCompareExchangePointer(mem: ptr pointer,
+#      newValue: pointer, comparand: pointer) : pointer {.nodecl, 
+#        importc: "InterlockedCompareExchangePointer", header:"windows.h".}
+
+#    proc compareAndSwap*[T](mem: ptr T, 
+#      expected: T, newValue: T): bool {.inline.}=
+#      ## Returns true if successfully set value at mem to newValue when value
+#      ## at mem == expected
+#      return InterlockedCompareExchangePointer(addr(mem), 
+#        addr(newValue), addr(expected))[] == expected
+    
+#elif not hasThreadSupport:
+#  proc compareAndSwap*[T](mem: ptr T, 
+#    expected: T, newValue: T): bool {.inline.} =
+#      ## Returns true if successfully set value at mem to newValue when value
+#      ## at mem == expected
+#      var oldval = mem[]
+#      if oldval == expected:
+#        mem[] = newValue
+#        return true
+#      return false
+
 
-proc atomicInc(memLoc: var int, x: int = 1): int =
-  when hasThreadSupport:
-    result = sync_add_and_fetch(memLoc, x)
+# Some convenient functions 
+proc atomicInc*(memLoc: var int, x: int = 1): int =
+  when defined(gcc) and hasThreadSupport:
+    result = atomic_add_fetch(memLoc.addr, x, ATOMIC_RELAXED)
   else:
     inc(memLoc, x)
     result = memLoc
   
-proc atomicDec(memLoc: var int, x: int = 1): int =
-  when hasThreadSupport:
-    when defined(sync_sub_and_fetch):
-      result = sync_sub_and_fetch(memLoc, x)
+proc atomicDec*(memLoc: var int, x: int = 1): int =
+  when defined(gcc) and hasThreadSupport:
+    when defined(atomic_sub_fetch):
+      result = atomic_sub_fetch(memLoc.addr, x, ATOMIC_RELAXED)
     else:
-      result = sync_add_and_fetch(memLoc, -x)
+      result = atomic_add_fetch(memLoc.addr, -x, ATOMIC_RELAXED)
   else:
     dec(memLoc, x)
     result = memLoc  
 
 
-# atomic compare and swap (CAS) funcitons to implement lock-free algorithms
-
-when (defined(gcc) or defined(llvm_gcc)) and hasThreadSupport:
-  proc compareAndSwap*[T: ptr|ref|pointer](mem: var T, expected: T, newValue: T): bool {.nodecl, 
-      importc: " __sync_bool_compare_and_swap".}
-    ## Returns true if successfully set value at mem to newValue when value
-    ## at mem == expected
-      
-elif defined(windows) and hasThreadSupport:
-    proc InterlockedCompareExchangePointer(mem: ptr pointer,
-      newValue: pointer, comparand: pointer) : pointer {.nodecl, 
-        importc: "InterlockedCompareExchangePointer", header:"windows.h".}
-
-
-    proc compareAndSwap*[T: ptr|ref|pointer](mem: var T, 
-      expected: T, newValue: T): bool {.inline.}=
-      ## Returns true if successfully set value at mem to newValue when value
-      ## at mem == expected
-      return InterlockedCompareExchangePointer(addr(mem), 
-        newValue, expected) == expected
-    
-elif not hasThreadSupport:
-  proc compareAndSwap*[T: ptr|ref|pointer](mem: var T, 
-    expected: T, newValue: T): bool {.inline.} =
-      ## Returns true if successfully set value at mem to newValue when value
-      ## at mem == expected
-      var oldval = mem
-      if oldval == expected:
-        mem = newValue
-        return true
-      return false
-