move threads out of system (#20674)

* move syslocks first

* progress

* clean up

* go on

* clean up

* clean up

* add imports syslocks

* remove documentation

* public deallocOsPages

* fixes genode

* fixes more

* fixes boehmGC

* cover more cases

* fixes cyclic deps

* fixes genode

* cleanup

* unpublic fields

* cleanup

* clean up

15 files changed, 425 insertions, 391 deletions

diff --git a/lib/core/locks.nim b/lib/core/locks.nim
index 92967b9db..ad0bff44d 100644
--- a/lib/core/locks.nim
+++ b/lib/core/locks.nim
@@ -18,8 +18,7 @@ when not compileOption("threads") and not defined(nimdoc):
   when false: # fix #12330
     {.error: "Locks requires --threads:on option.".}
 
-const insideRLocksModule = false
-include "system/syslocks"
+import std/private/syslocks
 
 type
   Lock* = SysLock ## Nim lock; whether this is re-entrant
diff --git a/lib/core/rlocks.nim b/lib/core/rlocks.nim
index 0444b9a83..bee5c1655 100644
--- a/lib/core/rlocks.nim
+++ b/lib/core/rlocks.nim
@@ -16,8 +16,7 @@ when not compileOption("threads") and not defined(nimdoc):
     # so they can replace each other seamlessly.
     {.error: "Rlocks requires --threads:on option.".}
 
-const insideRLocksModule = true
-include "system/syslocks"
+import std/private/syslocks
 
 type
   RLock* = SysLock ## Nim lock, re-entrant
diff --git a/lib/genode/alloc.nim b/lib/genode/alloc.nim
index 3ddd3074b..efc25ac57 100644
--- a/lib/genode/alloc.nim
+++ b/lib/genode/alloc.nim
@@ -15,7 +15,7 @@ when not defined(genode):
   {.error: "Genode only module".}
 
 when not declared(GenodeEnv):
-  include genode/env
+  import genode/env
 
 type RamDataspaceCapability {.
   importcpp: "Genode::Ram_dataspace_capability", pure.} = object
diff --git a/lib/genode/env.nim b/lib/genode/env.nim
index ef4a25883..34abf576f 100644
--- a/lib/genode/env.nim
+++ b/lib/genode/env.nim
@@ -20,10 +20,10 @@ when not defined(genode):
   {.error: "Genode only include".}
 
 type
-  GenodeEnvObj {.importcpp: "Genode::Env", header: "<base/env.h>", pure.} = object
-  GenodeEnvPtr = ptr GenodeEnvObj
+  GenodeEnvObj* {.importcpp: "Genode::Env", header: "<base/env.h>", pure.} = object
+  GenodeEnvPtr* = ptr GenodeEnvObj
 
-const runtimeEnvSym = "nim_runtime_env"
+const runtimeEnvSym* = "nim_runtime_env"
 
 when not defined(nimscript):
-  var runtimeEnv {.importcpp: runtimeEnvSym.}: GenodeEnvPtr
+  var runtimeEnv* {.importcpp: runtimeEnvSym.}: GenodeEnvPtr
diff --git a/lib/pure/concurrency/cpuinfo.nim b/lib/pure/concurrency/cpuinfo.nim
index 16d32002d..1d2ff63e1 100644
--- a/lib/pure/concurrency/cpuinfo.nim
+++ b/lib/pure/concurrency/cpuinfo.nim
@@ -40,7 +40,7 @@ when defined(macosx) or defined(bsd):
               importc: "sysctl", nodecl.}
 
 when defined(genode):
-  include genode/env
+  import genode/env
 
   proc affinitySpaceTotal(env: GenodeEnvPtr): cuint {.
     importcpp: "@->cpu().affinity_space().total()".}
diff --git a/lib/system/syslocks.nim b/lib/std/private/syslocks.nim
index 2f0c8b0ba..dff339e9e 100644
--- a/lib/system/syslocks.nim
+++ b/lib/std/private/syslocks.nim
@@ -15,7 +15,7 @@ when defined(windows):
   type
     Handle = int
 
-    SysLock {.importc: "CRITICAL_SECTION",
+    SysLock* {.importc: "CRITICAL_SECTION",
               header: "<windows.h>", final, pure.} = object # CRITICAL_SECTION in WinApi
       DebugInfo: pointer
       LockCount: int32
@@ -24,10 +24,10 @@ when defined(windows):
       LockSemaphore: int
       SpinCount: int
 
-    SysCond {.importc: "RTL_CONDITION_VARIABLE", header: "<windows.h>".} = object
+    SysCond* {.importc: "RTL_CONDITION_VARIABLE", header: "<windows.h>".} = object
       thePtr {.importc: "ptr".} : Handle
 
-  proc initSysLock(L: var SysLock) {.importc: "InitializeCriticalSection",
+  proc initSysLock*(L: var SysLock) {.importc: "InitializeCriticalSection",
                                      header: "<windows.h>".}
     ## Initializes the lock `L`.
 
@@ -35,18 +35,18 @@ when defined(windows):
                                                  header: "<windows.h>".}
     ## Tries to acquire the lock `L`.
 
-  proc tryAcquireSys(L: var SysLock): bool {.inline.} =
+  proc tryAcquireSys*(L: var SysLock): bool {.inline.} =
     result = tryAcquireSysAux(L) != 0'i32
 
-  proc acquireSys(L: var SysLock) {.importc: "EnterCriticalSection",
+  proc acquireSys*(L: var SysLock) {.importc: "EnterCriticalSection",
                                     header: "<windows.h>".}
     ## Acquires the lock `L`.
 
-  proc releaseSys(L: var SysLock) {.importc: "LeaveCriticalSection",
+  proc releaseSys*(L: var SysLock) {.importc: "LeaveCriticalSection",
                                     header: "<windows.h>".}
     ## Releases the lock `L`.
 
-  proc deinitSys(L: var SysLock) {.importc: "DeleteCriticalSection",
+  proc deinitSys*(L: var SysLock) {.importc: "DeleteCriticalSection",
                                    header: "<windows.h>".}
 
   proc initializeConditionVariable(
@@ -60,41 +60,41 @@ when defined(windows):
   ): int32 {.stdcall, noSideEffect, dynlib: "kernel32", importc: "SleepConditionVariableCS".}
 
 
-  proc signalSysCond(hEvent: var SysCond) {.stdcall, noSideEffect,
+  proc signalSysCond*(hEvent: var SysCond) {.stdcall, noSideEffect,
     dynlib: "kernel32", importc: "WakeConditionVariable".}
 
-  proc broadcastSysCond(hEvent: var SysCond) {.stdcall, noSideEffect,
+  proc broadcastSysCond*(hEvent: var SysCond) {.stdcall, noSideEffect,
     dynlib: "kernel32", importc: "WakeAllConditionVariable".}
 
-  proc initSysCond(cond: var SysCond) {.inline.} =
+  proc initSysCond*(cond: var SysCond) {.inline.} =
     initializeConditionVariable(cond)
-  proc deinitSysCond(cond: var SysCond) {.inline.} =
+  proc deinitSysCond*(cond: var SysCond) {.inline.} =
     discard
-  proc waitSysCond(cond: var SysCond, lock: var SysLock) =
+  proc waitSysCond*(cond: var SysCond, lock: var SysLock) =
     discard sleepConditionVariableCS(cond, lock, -1'i32)
 
 elif defined(genode):
   const
     Header = "genode_cpp/syslocks.h"
   type
-    SysLock {.importcpp: "Nim::SysLock", pure, final,
+    SysLock* {.importcpp: "Nim::SysLock", pure, final,
               header: Header.} = object
-    SysCond {.importcpp: "Nim::SysCond", pure, final,
+    SysCond* {.importcpp: "Nim::SysCond", pure, final,
               header: Header.} = object
 
-  proc initSysLock(L: var SysLock) = discard
-  proc deinitSys(L: var SysLock) = discard
-  proc acquireSys(L: var SysLock) {.noSideEffect, importcpp.}
-  proc tryAcquireSys(L: var SysLock): bool {.noSideEffect, importcpp.}
-  proc releaseSys(L: var SysLock) {.noSideEffect, importcpp.}
+  proc initSysLock*(L: var SysLock) = discard
+  proc deinitSys*(L: var SysLock) = discard
+  proc acquireSys*(L: var SysLock) {.noSideEffect, importcpp.}
+  proc tryAcquireSys*(L: var SysLock): bool {.noSideEffect, importcpp.}
+  proc releaseSys*(L: var SysLock) {.noSideEffect, importcpp.}
 
-  proc initSysCond(L: var SysCond) = discard
-  proc deinitSysCond(L: var SysCond) = discard
-  proc waitSysCond(cond: var SysCond, lock: var SysLock) {.
+  proc initSysCond*(L: var SysCond) = discard
+  proc deinitSysCond*(L: var SysCond) = discard
+  proc waitSysCond*(cond: var SysCond, lock: var SysLock) {.
     noSideEffect, importcpp.}
-  proc signalSysCond(cond: var SysCond) {.
+  proc signalSysCond*(cond: var SysCond) {.
     noSideEffect, importcpp.}
-  proc broadcastSysCond(cond: var SysCond) {.
+  proc broadcastSysCond*(cond: var SysCond) {.
     noSideEffect, importcpp.}
 
 else:
@@ -105,7 +105,7 @@ else:
       when defined(linux) and defined(amd64):
         abi: array[40 div sizeof(clong), clong]
 
-    SysLockAttr {.importc: "pthread_mutexattr_t", pure, final
+    SysLockAttr* {.importc: "pthread_mutexattr_t", pure, final
                header: """#include <sys/types.h>
                           #include <pthread.h>""".} = object
       when defined(linux) and defined(amd64):
@@ -143,8 +143,8 @@ else:
     # to prevent this once and for all, we're doing an extra malloc when
     # initializing the primitive.
     type
-      SysLock = ptr SysLockObj
-      SysCond = ptr SysCondObj
+      SysLock* = ptr SysLockObj
+      SysCond* = ptr SysCondObj
 
     when not declared(c_malloc):
       proc c_malloc(size: csize_t): pointer {.
@@ -152,83 +152,83 @@ else:
       proc c_free(p: pointer) {.
         importc: "free", header: "<stdlib.h>".}
 
-    proc initSysLock(L: var SysLock, attr: ptr SysLockAttr = nil) =
+    proc initSysLock*(L: var SysLock, attr: ptr SysLockAttr = nil) =
       L = cast[SysLock](c_malloc(csize_t(sizeof(SysLockObj))))
       initSysLockAux(L[], attr)
 
-    proc deinitSys(L: var SysLock) =
+    proc deinitSys*(L: var SysLock) =
       deinitSysAux(L[])
       c_free(L)
 
-    template acquireSys(L: var SysLock) =
+    template acquireSys*(L: var SysLock) =
       acquireSysAux(L[])
-    template tryAcquireSys(L: var SysLock): bool =
+    template tryAcquireSys*(L: var SysLock): bool =
       tryAcquireSysAux(L[]) == 0'i32
-    template releaseSys(L: var SysLock) =
+    template releaseSys*(L: var SysLock) =
       releaseSysAux(L[])
   else:
     type
-      SysLock = SysLockObj
-      SysCond = SysCondObj
+      SysLock* = SysLockObj
+      SysCond* = SysCondObj
 
-    template initSysLock(L: var SysLock, attr: ptr SysLockAttr = nil) =
+    template initSysLock*(L: var SysLock, attr: ptr SysLockAttr = nil) =
       initSysLockAux(L, attr)
-    template deinitSys(L: var SysLock) =
+    template deinitSys*(L: var SysLock) =
       deinitSysAux(L)
-    template acquireSys(L: var SysLock) =
+    template acquireSys*(L: var SysLock) =
       acquireSysAux(L)
-    template tryAcquireSys(L: var SysLock): bool =
+    template tryAcquireSys*(L: var SysLock): bool =
       tryAcquireSysAux(L) == 0'i32
-    template releaseSys(L: var SysLock) =
+    template releaseSys*(L: var SysLock) =
       releaseSysAux(L)
 
-  when insideRLocksModule:
-    let SysLockType_Reentrant {.importc: "PTHREAD_MUTEX_RECURSIVE",
-      header: "<pthread.h>".}: SysLockType
-    proc initSysLockAttr(a: var SysLockAttr) {.
-      importc: "pthread_mutexattr_init", header: "<pthread.h>", noSideEffect.}
-    proc setSysLockType(a: var SysLockAttr, t: SysLockType) {.
-      importc: "pthread_mutexattr_settype", header: "<pthread.h>", noSideEffect.}
+  # rlocks
+  var SysLockType_Reentrant* {.importc: "PTHREAD_MUTEX_RECURSIVE",
+    header: "<pthread.h>".}: SysLockType
+  proc initSysLockAttr*(a: var SysLockAttr) {.
+    importc: "pthread_mutexattr_init", header: "<pthread.h>", noSideEffect.}
+  proc setSysLockType*(a: var SysLockAttr, t: SysLockType) {.
+    importc: "pthread_mutexattr_settype", header: "<pthread.h>", noSideEffect.}
+
+  # locks
+  proc initSysCondAux(cond: var SysCondObj, cond_attr: ptr SysCondAttr = nil) {.
+    importc: "pthread_cond_init", header: "<pthread.h>", noSideEffect.}
+  proc deinitSysCondAux(cond: var SysCondObj) {.noSideEffect,
+    importc: "pthread_cond_destroy", header: "<pthread.h>".}
+
+  proc waitSysCondAux(cond: var SysCondObj, lock: var SysLockObj): cint {.
+    importc: "pthread_cond_wait", header: "<pthread.h>", noSideEffect.}
+  proc signalSysCondAux(cond: var SysCondObj) {.
+    importc: "pthread_cond_signal", header: "<pthread.h>", noSideEffect.}
+  proc broadcastSysCondAux(cond: var SysCondObj) {.
+    importc: "pthread_cond_broadcast", header: "<pthread.h>", noSideEffect.}
 
+  when defined(ios):
+    proc initSysCond*(cond: var SysCond, cond_attr: ptr SysCondAttr = nil) =
+      cond = cast[SysCond](c_malloc(csize_t(sizeof(SysCondObj))))
+      initSysCondAux(cond[], cond_attr)
+
+    proc deinitSysCond*(cond: var SysCond) =
+      deinitSysCondAux(cond[])
+      c_free(cond)
+
+    template waitSysCond*(cond: var SysCond, lock: var SysLock) =
+      discard waitSysCondAux(cond[], lock[])
+    template signalSysCond*(cond: var SysCond) =
+      signalSysCondAux(cond[])
+    template broadcastSysCond*(cond: var SysCond) =
+      broadcastSysCondAux(cond[])
   else:
-    proc initSysCondAux(cond: var SysCondObj, cond_attr: ptr SysCondAttr = nil) {.
-      importc: "pthread_cond_init", header: "<pthread.h>", noSideEffect.}
-    proc deinitSysCondAux(cond: var SysCondObj) {.noSideEffect,
-      importc: "pthread_cond_destroy", header: "<pthread.h>".}
-
-    proc waitSysCondAux(cond: var SysCondObj, lock: var SysLockObj): cint {.
-      importc: "pthread_cond_wait", header: "<pthread.h>", noSideEffect.}
-    proc signalSysCondAux(cond: var SysCondObj) {.
-      importc: "pthread_cond_signal", header: "<pthread.h>", noSideEffect.}
-    proc broadcastSysCondAux(cond: var SysCondObj) {.
-      importc: "pthread_cond_broadcast", header: "<pthread.h>", noSideEffect.}
-
-    when defined(ios):
-      proc initSysCond(cond: var SysCond, cond_attr: ptr SysCondAttr = nil) =
-        cond = cast[SysCond](c_malloc(csize_t(sizeof(SysCondObj))))
-        initSysCondAux(cond[], cond_attr)
-
-      proc deinitSysCond(cond: var SysCond) =
-        deinitSysCondAux(cond[])
-        c_free(cond)
-
-      template waitSysCond(cond: var SysCond, lock: var SysLock) =
-        discard waitSysCondAux(cond[], lock[])
-      template signalSysCond(cond: var SysCond) =
-        signalSysCondAux(cond[])
-      template broadcastSysCond(cond: var SysCond) =
-        broadcastSysCondAux(cond[])
-    else:
-      template initSysCond(cond: var SysCond, cond_attr: ptr SysCondAttr = nil) =
-        initSysCondAux(cond, cond_attr)
-      template deinitSysCond(cond: var SysCond) =
-        deinitSysCondAux(cond)
-
-      template waitSysCond(cond: var SysCond, lock: var SysLock) =
-        discard waitSysCondAux(cond, lock)
-      template signalSysCond(cond: var SysCond) =
-        signalSysCondAux(cond)
-      template broadcastSysCond(cond: var SysCond) =
-        broadcastSysCondAux(cond)
+    template initSysCond*(cond: var SysCond, cond_attr: ptr SysCondAttr = nil) =
+      initSysCondAux(cond, cond_attr)
+    template deinitSysCond*(cond: var SysCond) =
+      deinitSysCondAux(cond)
+
+    template waitSysCond*(cond: var SysCond, lock: var SysLock) =
+      discard waitSysCondAux(cond, lock)
+    template signalSysCond*(cond: var SysCond) =
+      signalSysCondAux(cond)
+    template broadcastSysCond*(cond: var SysCond) =
+      broadcastSysCondAux(cond)
 
 {.pop.}
diff --git a/lib/std/private/threadtypes.nim b/lib/std/private/threadtypes.nim
new file mode 100644
index 000000000..a1cdf21dc
--- /dev/null
+++ b/lib/std/private/threadtypes.nim
@@ -0,0 +1,176 @@
+include system/inclrtl
+
+const hasSharedHeap* = defined(boehmgc) or defined(gogc) # don't share heaps; every thread has its own
+
+when defined(windows):
+  type
+    Handle* = int
+    SysThread* = Handle
+    WinThreadProc* = proc (x: pointer): int32 {.stdcall.}
+
+  proc createThread*(lpThreadAttributes: pointer, dwStackSize: int32,
+                     lpStartAddress: WinThreadProc,
+                     lpParameter: pointer,
+                     dwCreationFlags: int32,
+                     lpThreadId: var int32): SysThread {.
+    stdcall, dynlib: "kernel32", importc: "CreateThread".}
+
+  proc winSuspendThread*(hThread: SysThread): int32 {.
+    stdcall, dynlib: "kernel32", importc: "SuspendThread".}
+
+  proc winResumeThread*(hThread: SysThread): int32 {.
+    stdcall, dynlib: "kernel32", importc: "ResumeThread".}
+
+  proc waitForSingleObject*(hHandle: SysThread, dwMilliseconds: int32): int32 {.
+    stdcall, dynlib: "kernel32", importc: "WaitForSingleObject".}
+
+  proc waitForMultipleObjects*(nCount: int32,
+                              lpHandles: ptr SysThread,
+                              bWaitAll: int32,
+                              dwMilliseconds: int32): int32 {.
+    stdcall, dynlib: "kernel32", importc: "WaitForMultipleObjects".}
+
+  proc terminateThread*(hThread: SysThread, dwExitCode: int32): int32 {.
+    stdcall, dynlib: "kernel32", importc: "TerminateThread".}
+
+  proc setThreadAffinityMask*(hThread: SysThread, dwThreadAffinityMask: uint) {.
+    importc: "SetThreadAffinityMask", stdcall, header: "<windows.h>".}
+
+elif defined(genode):
+  const
+    GenodeHeader* = "genode_cpp/threads.h"
+  type
+    SysThread* {.importcpp: "Nim::SysThread",
+                 header: GenodeHeader, final, pure.} = object
+    GenodeThreadProc* = proc (x: pointer) {.noconv.}
+
+  proc initThread*(s: var SysThread,
+                  env: GenodeEnv,
+                  stackSize: culonglong,
+                  entry: GenodeThreadProc,
+                  arg: pointer,
+                  affinity: cuint) {.
+    importcpp: "#.initThread(@)".}
+
+
+else:
+  when not (defined(macosx) or defined(haiku)):
+    {.passl: "-pthread".}
+
+  when not defined(haiku):
+    {.passc: "-pthread".}
+
+  const
+    schedh = "#define _GNU_SOURCE\n#include <sched.h>"
+    pthreadh* = "#define _GNU_SOURCE\n#include <pthread.h>"
+
+  when not declared(Time):
+    when defined(linux):
+      type Time = clong
+    else:
+      type Time = int
+
+  when (defined(linux) or defined(nintendoswitch)) and defined(amd64):
+    type
+      SysThread* {.importc: "pthread_t",
+                  header: "<sys/types.h>" .} = distinct culong
+      Pthread_attr* {.importc: "pthread_attr_t",
+                    header: "<sys/types.h>".} = object
+        abi: array[56 div sizeof(clong), clong]
+  elif defined(openbsd) and defined(amd64):
+    type
+      SysThread* {.importc: "pthread_t", header: "<pthread.h>".} = object
+      Pthread_attr* {.importc: "pthread_attr_t",
+                       header: "<pthread.h>".} = object
+  else:
+    type
+      SysThread* {.importc: "pthread_t", header: "<sys/types.h>".} = int
+      Pthread_attr* {.importc: "pthread_attr_t",
+                       header: "<sys/types.h>".} = object
+  type
+    Timespec* {.importc: "struct timespec", header: "<time.h>".} = object
+      tv_sec*: Time
+      tv_nsec*: clong
+
+  proc pthread_attr_init*(a1: var Pthread_attr): cint {.
+    importc, header: pthreadh.}
+  proc pthread_attr_setstack*(a1: ptr Pthread_attr, a2: pointer, a3: int): cint {.
+    importc, header: pthreadh.}
+  proc pthread_attr_setstacksize*(a1: var Pthread_attr, a2: int): cint {.
+    importc, header: pthreadh.}
+  proc pthread_attr_destroy*(a1: var Pthread_attr): cint {.
+    importc, header: pthreadh.}
+
+  proc pthread_create*(a1: var SysThread, a2: var Pthread_attr,
+            a3: proc (x: pointer): pointer {.noconv.},
+            a4: pointer): cint {.importc: "pthread_create",
+            header: pthreadh.}
+  proc pthread_join*(a1: SysThread, a2: ptr pointer): cint {.
+    importc, header: pthreadh.}
+
+  proc pthread_cancel*(a1: SysThread): cint {.
+    importc: "pthread_cancel", header: pthreadh.}
+
+  type CpuSet* {.importc: "cpu_set_t", header: schedh.} = object
+     when defined(linux) and defined(amd64):
+       abi: array[1024 div (8 * sizeof(culong)), culong]
+
+  proc cpusetZero*(s: var CpuSet) {.importc: "CPU_ZERO", header: schedh.}
+  proc cpusetIncl*(cpu: cint; s: var CpuSet) {.
+    importc: "CPU_SET", header: schedh.}
+
+  when defined(android):
+    # libc of android doesn't implement pthread_setaffinity_np,
+    # it exposes pthread_gettid_np though, so we can use that in combination
+    # with sched_setaffinity to set the thread affinity.
+    type Pid* {.importc: "pid_t", header: "<sys/types.h>".} = int32 # From posix_other.nim
+
+    proc setAffinityTID*(tid: Pid; setsize: csize_t; s: var CpuSet) {.
+      importc: "sched_setaffinity", header: schedh.}
+
+    proc pthread_gettid_np*(thread: SysThread): Pid {.
+      importc: "pthread_gettid_np", header: pthreadh.}
+
+    proc setAffinity*(thread: SysThread; setsize: csize_t; s: var CpuSet) =
+      setAffinityTID(pthread_gettid_np(thread), setsize, s)
+  else:
+    proc setAffinity*(thread: SysThread; setsize: csize_t; s: var CpuSet) {.
+      importc: "pthread_setaffinity_np", header: pthreadh.}
+
+
+const
+  emulatedThreadVars* = compileOption("tlsEmulation")
+# we preallocate a fixed size for thread local storage, so that no heap
+# allocations are needed. Currently less than 16K are used on a 64bit machine.
+# We use `float` for proper alignment:
+const nimTlsSize {.intdefine.} = 16000
+type
+  ThreadLocalStorage* = array[0..(nimTlsSize div sizeof(float)), float]
+  PGcThread* = ptr GcThread
+  GcThread* {.pure, inheritable.} = object
+    when emulatedThreadVars:
+      tls*: ThreadLocalStorage
+    else:
+      nil
+    when hasSharedHeap:
+      next*, prev*: PGcThread
+      stackBottom*, stackTop*: pointer
+      stackSize*: int
+    else:
+      nil
+
+const hasAllocStack* = defined(zephyr) # maybe freertos too?
+
+type
+  Thread*[TArg] = object
+    core*: PGcThread
+    sys*: SysThread
+    when TArg is void:
+      dataFn*: proc () {.nimcall, gcsafe.}
+    else:
+      dataFn*: proc (m: TArg) {.nimcall, gcsafe.}
+      data*: TArg
+    when hasAllocStack:
+      rawStack*: pointer
+
+proc `=copy`*[TArg](x: var Thread[TArg], y: Thread[TArg]) {.error.}
diff --git a/lib/system/threads.nim b/lib/std/threads.nim
index 4e190e443..5726a5cdc 100644
--- a/lib/system/threads.nim
+++ b/lib/std/threads.nim
@@ -9,10 +9,6 @@
 
 ## Thread support for Nim.
 ##
-## **Note**: This is part of the system module. Do not import it directly.
-## To activate thread support you need to compile
-## with the `--threads:on`:option: command line switch.
-##
 ## Nim's memory model for threads is quite different from other common
 ## programming languages (C, Pascal): Each thread has its own
 ## (garbage collected) heap and sharing of memory is restricted. This helps
@@ -44,24 +40,17 @@
 ##
 ##  deinitLock(L)
 
-when not declared(ThisIsSystem):
-  {.error: "You must not import this module explicitly".}
+import std/private/[threadtypes]
+export Thread
+
+import system/ansi_c
 
 when defined(nimPreviewSlimSystem):
   import std/assertions
 
-const
-  hasAllocStack = defined(zephyr) # maybe freertos too?
+when defined(genode):
+  import genode/env
 
-when defined(gcDestructors):
-  proc allocThreadStorage(size: int): pointer =
-    result = c_malloc(csize_t size)
-    zeroMem(result, size)
-
-  proc deallocThreadStorage(p: pointer) = c_free(p)
-else:
-  template allocThreadStorage(size: untyped): untyped = allocShared0(size)
-  template deallocThreadStorage(p: pointer) = deallocShared(p)
 
 when hasAllocStack or defined(zephyr) or defined(freertos):
   const
@@ -81,6 +70,14 @@ else:
 
     ThreadStackSize = ThreadStackMask+1 - StackGuardSize
 
+
+when defined(gcDestructors):
+  proc allocThreadStorage(size: int): pointer =
+    result = c_malloc(csize_t size)
+    zeroMem(result, size)
+else:
+  template allocThreadStorage(size: untyped): untyped = allocShared0(size)
+
 #const globalsSlot = ThreadVarSlot(0)
 #sysAssert checkSlot.int == globalsSlot.int
 
@@ -90,43 +87,13 @@ when defined(zephyr):
   #include <pthread.h>
   """.}
 
-# create for the main thread. Note: do not insert this data into the list
-# of all threads; it's not to be stopped etc.
-when not defined(useNimRtl):
-  #when not defined(createNimRtl): initStackBottom()
-  when declared(initGC):
-    initGC()
-    when not emulatedThreadVars:
-      type ThreadType {.pure.} = enum
-        None = 0,
-        NimThread = 1,
-        ForeignThread = 2
-      var
-        threadType {.rtlThreadVar.}: ThreadType
-
-      threadType = ThreadType.NimThread
 
 # We jump through some hops here to ensure that Nim thread procs can have
 # the Nim calling convention. This is needed because thread procs are
 # ``stdcall`` on Windows and ``noconv`` on UNIX. Alternative would be to just
 # use ``stdcall`` since it is mapped to ``noconv`` on UNIX anyway.
 
-type
-  Thread*[TArg] = object
-    core: PGcThread
-    sys: SysThread
-    when TArg is void:
-      dataFn: proc () {.nimcall, gcsafe.}
-    else:
-      dataFn: proc (m: TArg) {.nimcall, gcsafe.}
-      data: TArg
-    when hasAllocStack:
-      rawStack: pointer
-
-proc `=copy`*[TArg](x: var Thread[TArg], y: Thread[TArg]) {.error.}
 
-var
-  threadDestructionHandlers {.rtlThreadVar.}: seq[proc () {.closure, gcsafe, raises: [].}]
 
 proc onThreadDestruction*(handler: proc () {.closure, gcsafe, raises: [].}) =
   ## Registers a *thread local* handler that is called at the thread's
@@ -135,101 +102,20 @@ proc onThreadDestruction*(handler: proc () {.closure, gcsafe, raises: [].}) =
   ## A thread is destructed when the `.thread` proc returns
   ## normally or when it raises an exception. Note that unhandled exceptions
   ## in a thread nevertheless cause the whole process to die.
-  threadDestructionHandlers.add handler
-
-template afterThreadRuns() =
-  for i in countdown(threadDestructionHandlers.len-1, 0):
-    threadDestructionHandlers[i]()
-
-when not defined(boehmgc) and not hasSharedHeap and not defined(gogc) and not defined(gcRegions):
-  proc deallocOsPages() {.rtl, raises: [].}
-
-proc threadTrouble() {.raises: [], gcsafe.}
-  ## defined in system/excpt.nim
-
-when defined(boehmgc):
-  type GCStackBaseProc = proc(sb: pointer, t: pointer) {.noconv.}
-  proc boehmGC_call_with_stack_base(sbp: GCStackBaseProc, p: pointer)
-    {.importc: "GC_call_with_stack_base", boehmGC.}
-  proc boehmGC_register_my_thread(sb: pointer)
-    {.importc: "GC_register_my_thread", boehmGC.}
-  proc boehmGC_unregister_my_thread()
-    {.importc: "GC_unregister_my_thread", boehmGC.}
-
-  proc threadProcWrapDispatch[TArg](sb: pointer, thrd: pointer) {.noconv, raises: [].} =
-    boehmGC_register_my_thread(sb)
-    try:
-      let thrd = cast[ptr Thread[TArg]](thrd)
-      when TArg is void:
-        thrd.dataFn()
-      else:
-        thrd.dataFn(thrd.data)
-    except:
-      threadTrouble()
-    finally:
-      afterThreadRuns()
-    boehmGC_unregister_my_thread()
-else:
-  proc threadProcWrapDispatch[TArg](thrd: ptr Thread[TArg]) {.raises: [].} =
-    try:
-      when TArg is void:
-        thrd.dataFn()
-      else:
-        when defined(nimV2):
-          thrd.dataFn(thrd.data)
-        else:
-          var x: TArg
-          deepCopy(x, thrd.data)
-          thrd.dataFn(x)
-    except:
-      threadTrouble()
-    finally:
-      afterThreadRuns()
-      when hasAllocStack:
-        deallocThreadStorage(thrd.rawStack)
-
-proc threadProcWrapStackFrame[TArg](thrd: ptr Thread[TArg]) {.raises: [].} =
-  when defined(boehmgc):
-    boehmGC_call_with_stack_base(threadProcWrapDispatch[TArg], thrd)
-  elif not defined(nogc) and not defined(gogc) and not defined(gcRegions) and not usesDestructors:
-    var p {.volatile.}: pointer
-    # init the GC for refc/markandsweep
-    nimGC_setStackBottom(addr(p))
-    initGC()
-    when declared(threadType):
-      threadType = ThreadType.NimThread
-    threadProcWrapDispatch[TArg](thrd)
-    when declared(deallocOsPages): deallocOsPages()
-  else:
-    threadProcWrapDispatch(thrd)
-
-template threadProcWrapperBody(closure: untyped): untyped =
-  var thrd = cast[ptr Thread[TArg]](closure)
-  var core = thrd.core
-  when declared(globalsSlot): threadVarSetValue(globalsSlot, thrd.core)
-  threadProcWrapStackFrame(thrd)
-  # Since an unhandled exception terminates the whole process (!), there is
-  # no need for a ``try finally`` here, nor would it be correct: The current
-  # exception is tried to be re-raised by the code-gen after the ``finally``!
-  # However this is doomed to fail, because we already unmapped every heap
-  # page!
-
-  # mark as not running anymore:
-  thrd.core = nil
-  thrd.dataFn = nil
-  deallocThreadStorage(cast[pointer](core))
+  nimThreadDestructionHandlers.add handler
+
 
 {.push stack_trace:off.}
 when defined(windows):
   proc threadProcWrapper[TArg](closure: pointer): int32 {.stdcall.} =
-    threadProcWrapperBody(closure)
+    nimThreadProcWrapperBody(closure)
     # implicitly return 0
 elif defined(genode):
   proc threadProcWrapper[TArg](closure: pointer) {.noconv.} =
-    threadProcWrapperBody(closure)
+    nimThreadProcWrapperBody(closure)
 else:
   proc threadProcWrapper[TArg](closure: pointer): pointer {.noconv.} =
-    threadProcWrapperBody(closure)
+    nimThreadProcWrapperBody(closure)
 {.pop.}
 
 proc running*[TArg](t: Thread[TArg]): bool {.inline.} =
@@ -387,4 +273,4 @@ proc createThread*(t: var Thread[void], tp: proc () {.thread, nimcall.}) =
   createThread[void](t, tp)
 
 when not defined(gcOrc):
-  include threadids
+  include system/threadids
diff --git a/lib/system.nim b/lib/system.nim
index c28479a44..ff2bff45f 100644
--- a/lib/system.nim
+++ b/lib/system.nim
@@ -1115,7 +1115,7 @@ elif defined(nimdoc):
   proc quit*(errorcode: int = QuitSuccess) {.magic: "Exit", noreturn.}
 
 elif defined(genode):
-  include genode/env
+  import genode/env
 
   var systemEnv {.exportc: runtimeEnvSym.}: GenodeEnvPtr
 
@@ -1609,8 +1609,7 @@ when notJSnotNims:
 {.push stackTrace: off.}
 
 when not defined(js) and hasThreadSupport and hostOS != "standalone":
-  const insideRLocksModule = false
-  include "system/syslocks"
+  import std/private/syslocks
   include "system/threadlocalstorage"
 
 when not defined(js) and defined(nimV2):
@@ -2092,7 +2091,12 @@ when not defined(js):
   when declared(initAllocator):
     initAllocator()
   when hasThreadSupport:
-    when hostOS != "standalone": include "system/threads"
+    when hostOS != "standalone":
+      include system/threadimpl
+
+      import std/threads
+      export threads
+
   elif not defined(nogc) and not defined(nimscript):
     when not defined(useNimRtl) and not defined(createNimRtl): initStackBottom()
     when declared(initGC): initGC()
diff --git a/lib/system/alloc.nim b/lib/system/alloc.nim
index 88f680500..5350c1db8 100644
--- a/lib/system/alloc.nim
+++ b/lib/system/alloc.nim
@@ -11,6 +11,7 @@
 {.push profiler:off.}
 
 include osalloc
+import std/private/syslocks
 
 template track(op, address, size) =
   when defined(memTracker):
diff --git a/lib/system/channels_builtin.nim b/lib/system/channels_builtin.nim
index 2ad589383..50d740b23 100644
--- a/lib/system/channels_builtin.nim
+++ b/lib/system/channels_builtin.nim
@@ -139,6 +139,8 @@
 when not declared(ThisIsSystem):
   {.error: "You must not import this module explicitly".}
 
+import std/private/syslocks
+
 type
   pbytes = ptr UncheckedArray[byte]
   RawChannel {.pure, final.} = object ## msg queue for a thread
diff --git a/lib/system/gc_regions.nim b/lib/system/gc_regions.nim
index 6ced04c99..d96de7eac 100644
--- a/lib/system/gc_regions.nim
+++ b/lib/system/gc_regions.nim
@@ -7,6 +7,7 @@
 #
 
 # "Stack GC" for embedded devices or ultra performance requirements.
+import std/private/syslocks
 
 when defined(memProfiler):
   proc nimProfile(requestedSize: int) {.benign.}
diff --git a/lib/system/threadimpl.nim b/lib/system/threadimpl.nim
new file mode 100644
index 000000000..73f2807d2
--- /dev/null
+++ b/lib/system/threadimpl.nim
@@ -0,0 +1,107 @@
+var
+  nimThreadDestructionHandlers* {.rtlThreadVar.}: seq[proc () {.closure, gcsafe, raises: [].}]
+when not defined(boehmgc) and not hasSharedHeap and not defined(gogc) and not defined(gcRegions):
+  proc deallocOsPages() {.rtl, raises: [].}
+proc threadTrouble() {.raises: [], gcsafe.}
+# create for the main thread. Note: do not insert this data into the list
+# of all threads; it's not to be stopped etc.
+when not defined(useNimRtl):
+  #when not defined(createNimRtl): initStackBottom()
+  when declared(initGC):
+    initGC()
+    when not emulatedThreadVars:
+      type ThreadType {.pure.} = enum
+        None = 0,
+        NimThread = 1,
+        ForeignThread = 2
+      var
+        threadType {.rtlThreadVar.}: ThreadType
+
+      threadType = ThreadType.NimThread
+
+when defined(gcDestructors):
+  proc allocThreadStorage(size: int): pointer =
+    result = c_malloc(csize_t size)
+    zeroMem(result, size)
+
+  proc deallocThreadStorage(p: pointer) = c_free(p)
+else:
+  template allocThreadStorage(size: untyped): untyped = allocShared0(size)
+  template deallocThreadStorage(p: pointer) = deallocShared(p)
+
+template afterThreadRuns() =
+  for i in countdown(nimThreadDestructionHandlers.len-1, 0):
+    nimThreadDestructionHandlers[i]()
+
+when defined(boehmgc):
+  type GCStackBaseProc = proc(sb: pointer, t: pointer) {.noconv.}
+  proc boehmGC_call_with_stack_base(sbp: GCStackBaseProc, p: pointer)
+    {.importc: "GC_call_with_stack_base", boehmGC.}
+  proc boehmGC_register_my_thread(sb: pointer)
+    {.importc: "GC_register_my_thread", boehmGC.}
+  proc boehmGC_unregister_my_thread()
+    {.importc: "GC_unregister_my_thread", boehmGC.}
+
+  proc threadProcWrapDispatch[TArg](sb: pointer, thrd: pointer) {.noconv, raises: [].} =
+    boehmGC_register_my_thread(sb)
+    try:
+      let thrd = cast[ptr Thread[TArg]](thrd)
+      when TArg is void:
+        thrd.dataFn()
+      else:
+        thrd.dataFn(thrd.data)
+    except:
+      threadTrouble()
+    finally:
+      afterThreadRuns()
+    boehmGC_unregister_my_thread()
+else:
+  proc threadProcWrapDispatch[TArg](thrd: ptr Thread[TArg]) {.raises: [].} =
+    try:
+      when TArg is void:
+        thrd.dataFn()
+      else:
+        when defined(nimV2):
+          thrd.dataFn(thrd.data)
+        else:
+          var x: TArg
+          deepCopy(x, thrd.data)
+          thrd.dataFn(x)
+    except:
+      threadTrouble()
+    finally:
+      afterThreadRuns()
+      when hasAllocStack:
+        deallocThreadStorage(thrd.rawStack)
+
+proc threadProcWrapStackFrame[TArg](thrd: ptr Thread[TArg]) {.raises: [].} =
+  when defined(boehmgc):
+    boehmGC_call_with_stack_base(threadProcWrapDispatch[TArg], thrd)
+  elif not defined(nogc) and not defined(gogc) and not defined(gcRegions) and not usesDestructors:
+    var p {.volatile.}: pointer
+    # init the GC for refc/markandsweep
+    nimGC_setStackBottom(addr(p))
+    when declared(initGC):
+      initGC()
+    when declared(threadType):
+      threadType = ThreadType.NimThread
+    threadProcWrapDispatch[TArg](thrd)
+    when declared(deallocOsPages): deallocOsPages()
+  else:
+    threadProcWrapDispatch(thrd)
+
+template nimThreadProcWrapperBody*(closure: untyped): untyped =
+  var thrd = cast[ptr Thread[TArg]](closure)
+  var core = thrd.core
+  when declared(globalsSlot): threadVarSetValue(globalsSlot, thrd.core)
+  threadProcWrapStackFrame(thrd)
+  # Since an unhandled exception terminates the whole process (!), there is
+  # no need for a ``try finally`` here, nor would it be correct: The current
+  # exception is tried to be re-raised by the code-gen after the ``finally``!
+  # However this is doomed to fail, because we already unmapped every heap
+  # page!
+
+  # mark as not running anymore:
+  thrd.core = nil
+  thrd.dataFn = nil
+  deallocThreadStorage(cast[pointer](core))
diff --git a/lib/system/threadlocalstorage.nim b/lib/system/threadlocalstorage.nim
index b62c903c3..977f42e72 100644
--- a/lib/system/threadlocalstorage.nim
+++ b/lib/system/threadlocalstorage.nim
@@ -1,83 +1,33 @@
+import std/private/threadtypes
 
 when defined(windows):
   type
-    SysThread* = Handle
-    WinThreadProc = proc (x: pointer): int32 {.stdcall.}
-
-  proc createThread(lpThreadAttributes: pointer, dwStackSize: int32,
-                     lpStartAddress: WinThreadProc,
-                     lpParameter: pointer,
-                     dwCreationFlags: int32,
-                     lpThreadId: var int32): SysThread {.
-    stdcall, dynlib: "kernel32", importc: "CreateThread".}
-
-  proc winSuspendThread(hThread: SysThread): int32 {.
-    stdcall, dynlib: "kernel32", importc: "SuspendThread".}
-
-  proc winResumeThread(hThread: SysThread): int32 {.
-    stdcall, dynlib: "kernel32", importc: "ResumeThread".}
-
-  proc waitForSingleObject(hHandle: SysThread, dwMilliseconds: int32): int32 {.
-    stdcall, dynlib: "kernel32", importc: "WaitForSingleObject".}
-
-  proc waitForMultipleObjects(nCount: int32,
-                              lpHandles: ptr SysThread,
-                              bWaitAll: int32,
-                              dwMilliseconds: int32): int32 {.
-    stdcall, dynlib: "kernel32", importc: "WaitForMultipleObjects".}
-
-  proc terminateThread(hThread: SysThread, dwExitCode: int32): int32 {.
-    stdcall, dynlib: "kernel32", importc: "TerminateThread".}
-
-  type
     ThreadVarSlot = distinct int32
 
-  when true:
-    proc threadVarAlloc(): ThreadVarSlot {.
-      importc: "TlsAlloc", stdcall, header: "<windows.h>".}
-    proc threadVarSetValue(dwTlsIndex: ThreadVarSlot, lpTlsValue: pointer) {.
-      importc: "TlsSetValue", stdcall, header: "<windows.h>".}
-    proc tlsGetValue(dwTlsIndex: ThreadVarSlot): pointer {.
-      importc: "TlsGetValue", stdcall, header: "<windows.h>".}
+  proc threadVarAlloc(): ThreadVarSlot {.
+    importc: "TlsAlloc", stdcall, header: "<windows.h>".}
+  proc threadVarSetValue(dwTlsIndex: ThreadVarSlot, lpTlsValue: pointer) {.
+    importc: "TlsSetValue", stdcall, header: "<windows.h>".}
+  proc tlsGetValue(dwTlsIndex: ThreadVarSlot): pointer {.
+    importc: "TlsGetValue", stdcall, header: "<windows.h>".}
 
-    proc getLastError(): uint32 {.
-      importc: "GetLastError", stdcall, header: "<windows.h>".}
-    proc setLastError(x: uint32) {.
-      importc: "SetLastError", stdcall, header: "<windows.h>".}
+  proc getLastError(): uint32 {.
+    importc: "GetLastError", stdcall, header: "<windows.h>".}
+  proc setLastError(x: uint32) {.
+    importc: "SetLastError", stdcall, header: "<windows.h>".}
 
-    proc threadVarGetValue(dwTlsIndex: ThreadVarSlot): pointer =
-      let realLastError = getLastError()
-      result = tlsGetValue(dwTlsIndex)
-      setLastError(realLastError)
-  else:
-    proc threadVarAlloc(): ThreadVarSlot {.
-      importc: "TlsAlloc", stdcall, dynlib: "kernel32".}
-    proc threadVarSetValue(dwTlsIndex: ThreadVarSlot, lpTlsValue: pointer) {.
-      importc: "TlsSetValue", stdcall, dynlib: "kernel32".}
-    proc threadVarGetValue(dwTlsIndex: ThreadVarSlot): pointer {.
-      importc: "TlsGetValue", stdcall, dynlib: "kernel32".}
-
-  proc setThreadAffinityMask(hThread: SysThread, dwThreadAffinityMask: uint) {.
-    importc: "SetThreadAffinityMask", stdcall, header: "<windows.h>".}
+  proc threadVarGetValue(dwTlsIndex: ThreadVarSlot): pointer =
+    let realLastError = getLastError()
+    result = tlsGetValue(dwTlsIndex)
+    setLastError(realLastError)
 
 elif defined(genode):
-  import genode/env
   const
     GenodeHeader = "genode_cpp/threads.h"
+
   type
-    SysThread* {.importcpp: "Nim::SysThread",
-                 header: GenodeHeader, final, pure.} = object
-    GenodeThreadProc = proc (x: pointer) {.noconv.}
     ThreadVarSlot = int
 
-  proc initThread(s: var SysThread,
-                  env: GenodeEnv,
-                  stackSize: culonglong,
-                  entry: GenodeThreadProc,
-                  arg: pointer,
-                  affinity: cuint) {.
-    importcpp: "#.initThread(@)".}
-
   proc threadVarAlloc(): ThreadVarSlot = 0
 
   proc offMainThread(): bool {.
@@ -113,62 +63,18 @@ else:
   when not defined(haiku):
     {.passc: "-pthread".}
 
-  const
-    schedh = "#define _GNU_SOURCE\n#include <sched.h>"
-    pthreadh = "#define _GNU_SOURCE\n#include <pthread.h>"
-
-  when not declared(Time):
-    when defined(linux):
-      type Time = clong
-    else:
-      type Time = int
-
   when (defined(linux) or defined(nintendoswitch)) and defined(amd64):
     type
-      SysThread* {.importc: "pthread_t",
-                  header: "<sys/types.h>" .} = distinct culong
-      Pthread_attr {.importc: "pthread_attr_t",
-                    header: "<sys/types.h>".} = object
-        abi: array[56 div sizeof(clong), clong]
       ThreadVarSlot {.importc: "pthread_key_t",
                     header: "<sys/types.h>".} = distinct cuint
   elif defined(openbsd) and defined(amd64):
     type
-      SysThread* {.importc: "pthread_t", header: "<pthread.h>".} = object
-      Pthread_attr {.importc: "pthread_attr_t",
-                       header: "<pthread.h>".} = object
       ThreadVarSlot {.importc: "pthread_key_t",
                      header: "<pthread.h>".} = cint
   else:
     type
-      SysThread* {.importc: "pthread_t", header: "<sys/types.h>".} = int
-      Pthread_attr {.importc: "pthread_attr_t",
-                       header: "<sys/types.h>".} = object
       ThreadVarSlot {.importc: "pthread_key_t",
                      header: "<sys/types.h>".} = object
-  type
-    Timespec {.importc: "struct timespec", header: "<time.h>".} = object
-      tv_sec: Time
-      tv_nsec: clong
-
-  proc pthread_attr_init(a1: var Pthread_attr): cint {.
-    importc, header: pthreadh.}
-  proc pthread_attr_setstack*(a1: ptr Pthread_attr, a2: pointer, a3: int): cint {.
-    importc, header: pthreadh.}
-  proc pthread_attr_setstacksize(a1: var Pthread_attr, a2: int): cint {.
-    importc, header: pthreadh.}
-  proc pthread_attr_destroy(a1: var Pthread_attr): cint {.
-    importc, header: pthreadh.}
-
-  proc pthread_create(a1: var SysThread, a2: var Pthread_attr,
-            a3: proc (x: pointer): pointer {.noconv.},
-            a4: pointer): cint {.importc: "pthread_create",
-            header: pthreadh.}
-  proc pthread_join(a1: SysThread, a2: ptr pointer): cint {.
-    importc, header: pthreadh.}
-
-  proc pthread_cancel(a1: SysThread): cint {.
-    importc: "pthread_cancel", header: pthreadh.}
 
   proc pthread_getspecific(a1: ThreadVarSlot): pointer {.
     importc: "pthread_getspecific", header: pthreadh.}
@@ -188,59 +94,13 @@ else:
   proc threadVarGetValue(s: ThreadVarSlot): pointer {.inline.} =
     result = pthread_getspecific(s)
 
-  type CpuSet {.importc: "cpu_set_t", header: schedh.} = object
-     when defined(linux) and defined(amd64):
-       abi: array[1024 div (8 * sizeof(culong)), culong]
-
-  proc cpusetZero(s: var CpuSet) {.importc: "CPU_ZERO", header: schedh.}
-  proc cpusetIncl(cpu: cint; s: var CpuSet) {.
-    importc: "CPU_SET", header: schedh.}
-
-  when defined(android):
-    # libc of android doesn't implement pthread_setaffinity_np,
-    # it exposes pthread_gettid_np though, so we can use that in combination
-    # with sched_setaffinity to set the thread affinity.
-    type Pid {.importc: "pid_t", header: "<sys/types.h>".} = int32 # From posix_other.nim
-
-    proc setAffinityTID(tid: Pid; setsize: csize_t; s: var CpuSet) {.
-      importc: "sched_setaffinity", header: schedh.}
-
-    proc pthread_gettid_np(thread: SysThread): Pid {.
-      importc: "pthread_gettid_np", header: pthreadh.}
-
-    proc setAffinity(thread: SysThread; setsize: csize_t; s: var CpuSet) =
-      setAffinityTID(pthread_gettid_np(thread), setsize, s)
-  else:
-    proc setAffinity(thread: SysThread; setsize: csize_t; s: var CpuSet) {.
-      importc: "pthread_setaffinity_np", header: pthreadh.}
-
-
-const
-  emulatedThreadVars = compileOption("tlsEmulation")
 
 when emulatedThreadVars:
   # the compiler generates this proc for us, so that we can get the size of
   # the thread local var block; we use this only for sanity checking though
   proc nimThreadVarsSize(): int {.noconv, importc: "NimThreadVarsSize".}
 
-# we preallocate a fixed size for thread local storage, so that no heap
-# allocations are needed. Currently less than 16K are used on a 64bit machine.
-# We use `float` for proper alignment:
-const nimTlsSize {.intdefine.} = 16000
-type
-  ThreadLocalStorage = array[0..(nimTlsSize div sizeof(float)), float]
-  PGcThread = ptr GcThread
-  GcThread {.pure, inheritable.} = object
-    when emulatedThreadVars:
-      tls: ThreadLocalStorage
-    else:
-      nil
-    when hasSharedHeap:
-      next, prev: PGcThread
-      stackBottom, stackTop: pointer
-      stackSize: int
-    else:
-      nil
+
 
 when emulatedThreadVars:
   var globalsSlot: ThreadVarSlot
diff --git a/tools/kochdocs.nim b/tools/kochdocs.nim
index ee9c29499..c0bdb0503 100644
--- a/tools/kochdocs.nim
+++ b/tools/kochdocs.nim
@@ -138,7 +138,6 @@ mm.md
 """.splitWhitespace().mapIt("doc" / it)
 
   doc0 = """
-lib/system/threads.nim
 lib/system/channels_builtin.nim
 """.splitWhitespace() # ran by `nim doc0` instead of `nim doc`