thread support: next iteration

3 files changed, 107 insertions, 26 deletions

diff --git a/lib/system/excpt.nim b/lib/system/excpt.nim
index e2d3bea08..dac5678e0 100755
--- a/lib/system/excpt.nim
+++ b/lib/system/excpt.nim
@@ -81,8 +81,12 @@ when hasThreadSupport:
     proc pthread_setspecific(a1: Tpthread_key, a2: pointer): int32 {.
       importc: "pthread_setspecific", header: "<pthread.h>".}
     
-    proc specificDestroy(mem: pointer) {.noconv.} = dealloc(mem)
-    
+    proc specificDestroy(mem: pointer) {.noconv.} = 
+      #aquireSys(heapLock)
+      #dealloc(mem)
+      #releaseSys(heapLock)
+      #c_free(mem)
+
     proc ThreadVarAlloc(): TThreadVarSlot {.compilerproc, inline.} =
       discard pthread_key_create(addr(result), specificDestroy)
     proc ThreadVarSetValue(s: TThreadVarSlot, value: pointer) {.
@@ -104,31 +108,50 @@ when hasThreadSupport:
       tempFrames: array [0..127, PFrame] # cannot be allocated on the stack!
       data: float # compiler should add thread local variables here!
     PGlobals = ptr TGlobals
-
-  var globalsSlot = ThreadVarAlloc()
-  proc CreateThreadLocalStorage*(): pointer {.inl.} =
+  
+  # it's more efficient to not use a global variable for the thread storage 
+  # slot, but to rely on the implementation to assign slot 0 for us... ;-)
+  var checkSlot = ThreadVarAlloc()
+  const globalsSlot = TThreadVarSlot(0)
+  assert checkSlot.int == globalsSlot.int
+
+  proc AtomicAlloc0(size: int): pointer =
+    #AquireSys(heapLock)
+    result = c_malloc(size)
+    zeroMem(result, size)
+    #ReleaseSys(heapLock)
+
+  proc NewGlobals(): PGlobals = 
+    result = cast[PGlobals](AtomicAlloc0(sizeof(TGlobals)))
+    new(result.gAssertionFailed)
+    result.buf = newStringOfCap(2000)
+    result.assertBuf = newStringOfCap(2000)
+
+  proc AllocThreadLocalStorage*(): pointer {.inl.} =
     isMultiThreaded = true
-    result = alloc0(sizeof(TGlobals))
-    ThreadVarSetValue(globalsSlot, result)
+    result = NewGlobals()
+    
+  proc SetThreadLocalStorage*(p: pointer) {.inl.} =
+    ThreadVarSetValue(globalsSlot, p)
     
   proc GetGlobals(): PGlobals {.compilerRtl, inl.} =
     result = cast[PGlobals](ThreadVarGetValue(globalsSlot))
 
   # create for the main thread:
-  ThreadVarSetValue(globalsSlot, alloc0(sizeof(TGlobals)))
+  ThreadVarSetValue(globalsSlot, NewGlobals())
 
 when hasThreadSupport:
-  template ThreadGlobals = 
+  template ThreadGlobals =
     var globals = GetGlobals()
   template `||`(varname: expr): expr = globals.varname
   
-  ThreadGlobals()
+  #ThreadGlobals()
 else:
   template ThreadGlobals = nil # nothing
   template `||`(varname: expr): expr = varname
 
   var
-    framePtr {.compilerproc.}: PFrame # XXX only temporarily a compilerproc
+    framePtr: PFrame
     excHandler: PSafePoint = nil
       # list of exception handlers
       # a global variable for the root of all try blocks
@@ -141,6 +164,11 @@ else:
     tempFrames: array [0..127, PFrame] # cannot be allocated on the stack!
     gAssertionFailed: ref EAssertionFailed
 
+  new(||gAssertionFailed)
+  ||buf = newStringOfCap(2000)
+  ||assertBuf = newStringOfCap(2000)
+
+
 proc pushFrame(s: PFrame) {.compilerRtl, inl.} = 
   ThreadGlobals()
   s.prev = ||framePtr
@@ -388,11 +416,6 @@ proc registerSignalHandler() =
 
 when not defined(noSignalHandler):
   registerSignalHandler() # call it in initialization section
-# for easier debugging of the GC, this memory is only allocated after the
-# signal handlers have been registered
-new(||gAssertionFailed)
-||buf = newStringOfCap(2000)
-||assertBuf = newStringOfCap(2000)
 
 proc raiseRangeError(val: biggestInt) {.compilerproc, noreturn, noinline.} =
   raise newException(EOutOfRange, "value " & $val & " out of range")
diff --git a/lib/system/gc.nim b/lib/system/gc.nim
index 52de66d48..3a7270539 100755
--- a/lib/system/gc.nim
+++ b/lib/system/gc.nim
@@ -81,14 +81,14 @@ var
 proc aquire(gch: var TGcHeap) {.inline.} = 
   when hasThreadSupport:
     if isMultiThreaded: 
-      aquire(gch.zctLock)
-      aquire(gch.cycleRootsLock)
+      aquireSys(gch.zctLock)
+      aquireSys(gch.cycleRootsLock)
 
 proc release(gch: var TGcHeap) {.inline.} = 
   when hasThreadSupport:
     if isMultiThreaded: 
-      release(gch.zctLock)
-      release(gch.cycleRootsLock)
+      releaseSys(gch.zctLock)
+      releaseSys(gch.cycleRootsLock)
 
 proc addZCT(s: var TCellSeq, c: PCell) {.noinline.} =
   if (c.refcount and rcZct) == 0:
@@ -207,18 +207,18 @@ proc prepareDealloc(cell: PCell) =
 proc rtlAddCycleRoot(c: PCell) {.rtl, inl.} = 
   # we MUST access gch as a global here, because this crosses DLL boundaries!
   when hasThreadSupport:
-    if isMultiThreaded: Aquire(gch.cycleRootsLock)
+    if isMultiThreaded: AquireSys(gch.cycleRootsLock)
   incl(gch.cycleRoots, c)
   when hasThreadSupport:  
-    if isMultiThreaded: Release(gch.cycleRootsLock)
+    if isMultiThreaded: ReleaseSys(gch.cycleRootsLock)
 
 proc rtlAddZCT(c: PCell) {.rtl, inl.} =
   # we MUST access gch as a global here, because this crosses DLL boundaries!
   when hasThreadSupport:
-    if isMultiThreaded: Aquire(gch.zctLock)
+    if isMultiThreaded: AquireSys(gch.zctLock)
   addZCT(gch.zct, c)
   when hasThreadSupport:
-    if isMultiThreaded: Release(gch.zctLock)
+    if isMultiThreaded: ReleaseSys(gch.zctLock)
 
 proc decRef(c: PCell) {.inline.} =
   when stressGC:
@@ -287,9 +287,10 @@ proc initGC() =
     Init(gch.cycleRoots)
     Init(gch.decStack)
     when hasThreadSupport:
-      InitLock(gch.cycleRootsLock)
-      InitLock(gch.zctLock)
+      InitSysLock(gch.cycleRootsLock)
+      InitSysLock(gch.zctLock)
     new(gOutOfMem) # reserve space for the EOutOfMemory exception here!
+    
 
 proc forAllSlotsAux(dest: pointer, n: ptr TNimNode, op: TWalkOp) =
   var d = cast[TAddress](dest)
diff --git a/lib/system/systhread.nim b/lib/system/systhread.nim
index d9b340ce2..2b5057ff0 100755
--- a/lib/system/systhread.nim
+++ b/lib/system/systhread.nim
@@ -9,6 +9,12 @@
 
 const
   maxThreads = 256
+  SystemInclude = defined(hasThreadSupport)
+
+when not SystemInclude:
+  # ugly hack: this file is then included from core/threads, so we have
+  # thread support:
+  const hasThreadSupport = true
 
 when (defined(gcc) or defined(llvm_gcc)) and hasThreadSupport:
   proc sync_add_and_fetch(p: var int, val: int): int {.
@@ -40,3 +46,54 @@ proc atomicDec(memLoc: var int, x: int): int =
     dec(memLoc, x)
     result = memLoc  
 
+when defined(Windows):
+  type
+    TSysLock {.final, pure.} = object # CRITICAL_SECTION in WinApi
+      DebugInfo: pointer
+      LockCount: int32
+      RecursionCount: int32
+      OwningThread: int
+      LockSemaphore: int
+      Reserved: int32
+          
+  proc InitSysLock(L: var TSysLock) {.stdcall,
+    dynlib: "kernel32", importc: "InitializeCriticalSection".}
+    ## Initializes the lock `L`.
+
+  proc TryAquireSysAux(L: var TSysLock): int32 {.stdcall,
+    dynlib: "kernel32", importc: "TryEnterCriticalSection".}
+    ## Tries to aquire the lock `L`.
+    
+  proc TryAquireSys(L: var TSysLock): bool {.inline.} = 
+    result = TryAquireSysAux(L) != 0'i32
+
+  proc AquireSys(L: var TSysLock) {.stdcall,
+    dynlib: "kernel32", importc: "EnterCriticalSection".}
+    ## Aquires the lock `L`.
+    
+  proc ReleaseSys(L: var TSysLock) {.stdcall,
+    dynlib: "kernel32", importc: "LeaveCriticalSection".}
+    ## Releases the lock `L`.
+
+else:
+  type
+    TSysLock {.importc: "pthread_mutex_t", header: "<sys/types.h>".} = int
+
+  proc InitSysLock(L: var TSysLock, attr: pointer = nil) {.
+    importc: "pthread_mutex_init", header: "<pthread.h>".}
+
+  proc AquireSys(L: var TSysLock) {.
+    importc: "pthread_mutex_lock", header: "<pthread.h>".}
+  proc TryAquireSysAux(L: var TSysLock): cint {.
+    importc: "pthread_mutex_trylock", header: "<pthread.h>".}
+
+  proc TryAquireSys(L: var TSysLock): bool {.inline.} = 
+    result = TryAquireSysAux(L) == 0'i32
+
+  proc ReleaseSys(L: var TSysLock) {.
+    importc: "pthread_mutex_unlock", header: "<pthread.h>".}
+
+when SystemInclude:
+  var heapLock: TSysLock
+  InitSysLock(HeapLock)
+