---
 rtl/win/systlsdir.inc | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/rtl/win/systlsdir.inc b/rtl/win/systlsdir.inc
index 328f0b141a..af60fee15b 100644
--- a/rtl/win/systlsdir.inc
+++ b/rtl/win/systlsdir.inc
@@ -115,12 +115,17 @@ procedure Exec_Tls_callback(Handle : pointer; reason : Dword; Reserved : pointer
          end;
        DLL_THREAD_DETACH :
          begin
+         {$ifdef FPC_SECTION_THREADVARS}
            if ThreadID<>0 then
+         {$else !FPC_SECTION_THREADVARS}
+           if TlsGetValue(TLSKey^)<>nil then
+         {$endif !FPC_SECTION_THREADVARS}
              DoneThread; { Assume everything is idempotent there }
          end;
      end;
   end;
 
+
 { Mingw tlssup.c source code has
   _CRTALLOC(".CRT$XLA") PIMAGE_TLS_CALLBACK __xl_a = 0;
   _CRTALLOC(".CRT$XLZ") PIMAGE_TLS_CALLBACK __xl_z = 0;
-- 
2.26.2.windows.1

---
 rtl/win/systhrd.inc   | 8 +++++---
 rtl/win/systlsdir.inc | 2 +-
 2 files changed, 6 insertions(+), 4 deletions(-)

diff --git a/rtl/win/systhrd.inc b/rtl/win/systhrd.inc
index 1abc04b53e..dfa583fee7 100644
--- a/rtl/win/systhrd.inc
+++ b/rtl/win/systhrd.inc
@@ -19,6 +19,7 @@
                            Local WINApi imports
 *****************************************************************************}
 
+{$ifndef FPC_SECTION_THREADVARS}
 const
   { LocalAlloc flags  }
   LMEM_FIXED = 0;
@@ -34,15 +35,16 @@ function TlsGetValue(dwTlsIndex : DWord) : pointer;
   {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'TlsGetValue';
 function TlsSetValue(dwTlsIndex : DWord;lpTlsValue : pointer) : LongBool;
   {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'TlsSetValue';
+function LocalAlloc(uFlags:DWord; dwBytes:SIZE_T):Pointer;
+  {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'LocalAlloc';
+function LocalFree(hMem : Pointer):Pointer; {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'LocalFree';
+{$endif !FPC_SECTION_THREADVARS}
 function CreateThread(lpThreadAttributes : pointer;
   dwStackSize : SIZE_T; lpStartAddress : pointer;lpParameter : pointer;
   dwCreationFlags : DWord;var lpThreadId : DWord) : THandle;
   {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'CreateThread';
 procedure ExitThread(dwExitCode : DWord);
   {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'ExitThread';
-function LocalAlloc(uFlags:DWord; dwBytes:SIZE_T):Pointer;
-  {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'LocalAlloc';
-function LocalFree(hMem : Pointer):Pointer; {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'LocalFree';
 procedure Sleep(dwMilliseconds: DWord); {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'Sleep';
 function  WinSuspendThread (threadHandle : THandle) : dword; {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'SuspendThread';
 function  WinResumeThread  (threadHandle : THandle) : dword; {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'ResumeThread';
diff --git a/rtl/win/systlsdir.inc b/rtl/win/systlsdir.inc
index af60fee15b..28bd4e9be3 100644
--- a/rtl/win/systlsdir.inc
+++ b/rtl/win/systlsdir.inc
@@ -46,10 +46,10 @@ type
   end;
 
 
+{$ifndef FPC_SECTION_THREADVARS}
 function TlsGetValue(dwTlsIndex : DWord) : pointer; stdcall;
   external 'kernel32' name 'TlsGetValue';
 
-{$ifndef FPC_SECTION_THREADVARS}
 procedure InitSystemThreads; external name '_FPC_InitSystemThreads';
 procedure SysAllocateThreadVars; external name '_FPC_SysAllocateThreadVars';
 procedure InitHeap; external name '_FPC_InitHeap';
-- 
2.26.2.windows.1

---
 compiler/aggas.pas         |  4 +-
 compiler/cgbase.pas        |  1 +
 compiler/dbgdwarf.pas      |  3 +-
 compiler/nld.pas           |  4 +-
 compiler/nutils.pas        |  9 +++--
 compiler/ogcoff.pas        | 28 ++++----------
 compiler/systems/i_win.pas |  3 ++
 compiler/x86/aasmcpu.pas   |  6 +++
 compiler/x86/agx86att.pas  |  9 +++++
 compiler/x86/nx86ld.pas    | 79 +++++++++++++++++++++++---------------
 rtl/inc/heaptrc.pp         |  4 ++
 rtl/linux/si_impl.inc      |  2 +
 rtl/win/sysosh.inc         |  5 +--
 rtl/win/systhrd.inc        | 27 ++++++++++---
 rtl/win/systlsdir.inc      | 14 ++++++-
 rtl/win/syswin.inc         |  8 ++++
 rtl/win32/sysinit.inc      | 10 +++++
 rtl/win32/system.pp        | 37 +-----------------
 rtl/win64/sysinit.pp       | 16 ++++++++
 rtl/win64/system.pp        | 43 +--------------------
 20 files changed, 167 insertions(+), 145 deletions(-)

diff --git a/compiler/aggas.pas b/compiler/aggas.pas
index 41d20c5129..f9b69f605b 100644
--- a/compiler/aggas.pas
+++ b/compiler/aggas.pas
@@ -533,7 +533,9 @@ implementation
              { according to the GNU AS guide AS for COFF does not support the
                progbits }
              writer.AsmWrite('.section ');
-             usesectionflags:=true;
+             { .tls section requires special processing }
+             if not((atype=sec_threadvar) and (tf_section_threadvars in target_info.flags)) then
+               usesectionflags:=true;
            end;
          system_powerpc_darwin,
          system_i386_darwin,
diff --git a/compiler/cgbase.pas b/compiler/cgbase.pas
index 6dca71ba40..0e862ca739 100644
--- a/compiler/cgbase.pas
+++ b/compiler/cgbase.pas
@@ -140,6 +140,7 @@ interface
 {$ifdef x86_64}
           ,addr_tpoff
           ,addr_tlsgd
+          ,addr_secrel
 {$endif x86_64}
          );
 
diff --git a/compiler/dbgdwarf.pas b/compiler/dbgdwarf.pas
index 5b75e257b5..d81436ed2a 100644
--- a/compiler/dbgdwarf.pas
+++ b/compiler/dbgdwarf.pas
@@ -2658,7 +2658,8 @@ implementation
                   begin
                     if vo_is_thread_var in sym.varoptions then
                       begin
-                        if tf_section_threadvars in target_info.flags then
+                        if (tf_section_threadvars in target_info.flags) and
+                           (not (target_info.system in [system_x86_64_win64])) then // TODO : dwarf for threadvars on win64
                           begin
                             case sizeof(puint) of
                               2:
diff --git a/compiler/nld.pas b/compiler/nld.pas
index e2bccbc080..0453f162e8 100644
--- a/compiler/nld.pas
+++ b/compiler/nld.pas
@@ -450,7 +450,9 @@ implementation
                 { call to get address of threadvar }
                 if (vo_is_thread_var in tabstractvarsym(symtableentry).varoptions) then
                   begin
-                    include(current_procinfo.flags,pi_do_call);
+                    if (not (tf_section_threadvars in target_info.flags)) or
+                       (current_settings.tlsmodel in [tlsm_local_dynamic,tlsm_global_dynamic]) then
+                      include(current_procinfo.flags,pi_do_call);
                     include(current_procinfo.flags,pi_uses_threadvar);
                   end;
               end;
diff --git a/compiler/nutils.pas b/compiler/nutils.pas
index d586fc222d..b5c0de1937 100644
--- a/compiler/nutils.pas
+++ b/compiler/nutils.pas
@@ -191,7 +191,7 @@ interface
 implementation
 
     uses
-      cutils,verbose,globals,
+      cutils,verbose,globals,systems,
       symconst,symdef,
       defcmp,defutil,
       ncon,ncnv,nld,nflw,nset,ncal,nadd,nmem,ninl,
@@ -799,10 +799,13 @@ implementation
                 begin
                   if assigned(tloadnode(p).left) then
                     inc(result,node_complexity(tloadnode(p).left));
-                  { threadvars need a helper call }
+                  { some threadvars models need a helper call }
                   if (tloadnode(p).symtableentry.typ=staticvarsym) and
                      (vo_is_thread_var in tstaticvarsym(tloadnode(p).symtableentry).varoptions) then
-                    inc(result,5)
+                    if (tf_section_threadvars in target_info.flags) and (current_settings.tlsmodel in [tlsm_local_exec,tlsm_initial_exec]) then
+                      inc(result,3)
+                    else
+                      inc(result,50)
                   else if not((tloadnode(p).symtableentry.typ in [staticvarsym,localvarsym,paravarsym,fieldvarsym]) and
                     (tabstractvarsym(tloadnode(p).symtableentry).varregable in [vr_intreg,vr_mmreg,vr_fpureg])) then
                     inc(result);
diff --git a/compiler/ogcoff.pas b/compiler/ogcoff.pas
index 8c05b3792a..803b9ebad3 100644
--- a/compiler/ogcoff.pas
+++ b/compiler/ogcoff.pas
@@ -1015,8 +1015,10 @@ const pemagic : array[0..3] of byte = (
                   end;
                 RELOC_SECREL32 :
                   begin
-                    { fixup address when the symbol was known in defined object }
-                    if (relocsec.objdata=objsec.objdata) then
+                    { fixup address when the symbol was known in defined object or
+                      when the symbol is a threadvar }
+                    if (relocsec.objdata=objsec.objdata) or
+                       assigned(objreloc.symbol) and (copy(objreloc.symbol.objsection.Name,1,4)='.tls') then
                       dec(address,relocsec.ExeSection.MemPos);
                     inc(address,relocval);
                   end;
@@ -2667,21 +2669,17 @@ const pemagic : array[0..3] of byte = (
 
         procedure UpdateTlsDataDir;
         var
-          {callbacksection : TExeSection;}
           tlsexesymbol: TExeSymbol;
           tlssymbol: TObjSymbol;
-          callbackexesymbol: TExeSymbol;
-          //callbacksymbol: TObjSymbol;
         begin
           { according to GNU ld,
             the callback routines should be placed into .CRT$XL*
             sections, and the thread local variables in .tls
             __tls_start__ and __tls_end__ symbols
-            should be used for the initialized part,
-            which we do not support yet. }
-          { For now, we only pass the address of the __tls_used
+            should be used for the initialized part. }
+          { We pass the address of the __tls_used
             asm symbol into PE_DATADIR_TLS with the correct
-            size of this table (different for win32/win64 }
+            size of this table (different for win32/win64) }
           tlsexesymbol:=texesymbol(ExeSymbolList.Find(
             target_info.Cprefix+'_tls_used'));
           if assigned(tlsexesymbol) then
@@ -2690,18 +2688,6 @@ const pemagic : array[0..3] of byte = (
               peoptheader.DataDirectory[PE_DATADIR_TLS].vaddr:=tlssymbol.address;
               { sizeof(TlsDirectory) is different on host and target when cross-compiling }
               peoptheader.DataDirectory[PE_DATADIR_TLS].size:=TLSDIR_SIZE;
-              if IsSharedLibrary then
-                begin
-                  { Here we should reset __FPC_tls_callbacks value to nil }
-                  callbackexesymbol:=texesymbol(ExeSymbolList.Find(
-                                        '__FPC_tls_callbacks'));
-                  if assigned (callbackexesymbol) then
-                    begin
-                      //callbacksymbol:=callbackexesymbol.ObjSymbol;
-
-                    end;
-                end;
-
            end;
         end;
 
diff --git a/compiler/systems/i_win.pas b/compiler/systems/i_win.pas
index d4e8f842d5..d1ffb4e57c 100644
--- a/compiler/systems/i_win.pas
+++ b/compiler/systems/i_win.pas
@@ -107,6 +107,9 @@ unit i_win;
             shortname    : 'Win64';
             flags        : [tf_files_case_aware,tf_has_dllscanner,
                             tf_smartlink_sections,
+{$if not defined(DISABLE_TLS_DIRECTORY) and not defined(SUPPORT_WINXP)}
+                            tf_section_threadvars,
+{$endif}
                             tf_winlikewidestring,tf_no_pic_supported,
                             tf_dwarf_only_local_labels,
                             tf_no_generic_stackcheck,tf_has_winlike_resources,
diff --git a/compiler/x86/aasmcpu.pas b/compiler/x86/aasmcpu.pas
index f712af7196..115474e6d1 100644
--- a/compiler/x86/aasmcpu.pas
+++ b/compiler/x86/aasmcpu.pas
@@ -3617,6 +3617,12 @@ implementation
                       currabsreloc:=RELOC_TLSGD;
                       currabsreloc32:=RELOC_TLSGD;
                     end
+                  else if oper[opidx]^.ref^.refaddr=addr_secrel then
+                    begin
+                      currrelreloc:=RELOC_SECREL32;
+                      currabsreloc:=RELOC_SECREL32;
+                      currabsreloc32:=RELOC_SECREL32;
+                    end
                   else
 {$endif x86_64}
                     begin
diff --git a/compiler/x86/agx86att.pas b/compiler/x86/agx86att.pas
index 657f073163..bf69e08bd5 100644
--- a/compiler/x86/agx86att.pas
+++ b/compiler/x86/agx86att.pas
@@ -190,6 +190,8 @@ interface
                owner.writer.AsmWrite('@tpoff');
              addr_tlsgd:
                owner.writer.AsmWrite('@tlsgd');
+             addr_secrel:
+               owner.writer.AsmWrite('@secrel32');
 {$endif x86_64}
              else
                ;
@@ -259,6 +261,13 @@ interface
               {$ifdef x86_64},addr_tpoff,addr_tlsgd{$endif x86_64}
               ] then
               WriteReference(o.ref^)
+        {$ifdef x86_64}
+            else if o.ref^.refaddr=addr_secrel then
+              begin
+                owner.writer.AsmWrite('$');
+                WriteReference(o.ref^);
+              end
+        {$endif x86_64}
             else
               begin
                 owner.writer.AsmWrite('$');
diff --git a/compiler/x86/nx86ld.pas b/compiler/x86/nx86ld.pas
index 774addd1ca..61c82755d4 100644
--- a/compiler/x86/nx86ld.pas
+++ b/compiler/x86/nx86ld.pas
@@ -38,10 +38,10 @@ interface
 implementation
 
     uses
-      globals,
+      globals,fmodule,
       cutils,verbose,systems,
       aasmbase,aasmtai,aasmdata,aasmcpu,
-      cgutils,cgobj,
+      cgutils,cgobj,hlcgobj,
       symconst,symdef,symtable,
       cgbase,cpubase,parabase,paramgr,
       procinfo;
@@ -54,37 +54,10 @@ implementation
       var
         paraloc1 : tcgpara;
         pd: tprocdef;
-        href: treference;
-        hregister : tregister;
-        handled: boolean;
+        href, href_index: treference;
+        hregister, reg_index, reg_offset, reg_start : tregister;
       begin
-        handled:=false;
-        if (tf_section_threadvars in target_info.flags) then
-          begin
-            if target_info.system in [system_i386_win32,system_x86_64_win64] then
-              begin
-                paraloc1.init;
-                pd:=search_system_proc('fpc_tls_add');
-                paramanager.getcgtempparaloc(current_asmdata.CurrAsmList,pd,1,paraloc1);
-                if not(vo_is_weak_external in gvs.varoptions) then
-                  reference_reset_symbol(href,current_asmdata.RefAsmSymbol(gvs.mangledname,AT_DATA,use_indirect_symbol(gvs)),0,sizeof(pint),[])
-                else
-                  reference_reset_symbol(href,current_asmdata.WeakRefAsmSymbol(gvs.mangledname,AT_DATA),0,sizeof(pint),[]);
-                cg.a_loadaddr_ref_cgpara(current_asmdata.CurrAsmList,href,paraloc1);
-                paramanager.freecgpara(current_asmdata.CurrAsmList,paraloc1);
-                paraloc1.done;
-
-                cg.g_call(current_asmdata.CurrAsmList,'FPC_TLS_ADD');
-                cg.ungetcpuregister(current_asmdata.CurrAsmList,NR_FUNCTION_RESULT_REG);
-                hregister:=cg.getaddressregister(current_asmdata.CurrAsmList);
-                cg.a_load_reg_reg(current_asmdata.CurrAsmList,OS_ADDR,OS_ADDR,NR_FUNCTION_RESULT_REG,hregister);
-                location.reference.base:=hregister;
-                handled:=true;
-              end;
-          end;
-
-        if not handled then
-          inherited;
+        inherited;
 
         if (tf_section_threadvars in target_info.flags) then
           begin
@@ -156,6 +129,48 @@ implementation
                       Internalerror(2019012002);
                   end;
                 end;
+              system_x86_64_win64:
+                begin
+                  reference_reset(href,sizeof(AInt),[]); href.segment:=NR_GS; href.offset:=88;
+                  hregister:=cg.getaddressregister(current_asmdata.CurrAsmList);
+                  current_asmdata.CurrAsmList.concat(taicpu.op_ref_reg(A_MOV,S_Q,href,hregister));         // movq %gs:88, %rcx
+
+                  reference_reset_base(href,hregister,0,ctempposinvalid,sizeof(AInt),[]);
+                  if (main_module=nil) or main_module.IsLibrary or main_module.is_unit then
+                    begin // _tls_index <> 0
+                      reference_reset_symbol(href_index,current_asmdata.RefAsmSymbol('_tls_index',AT_DATA),0,sizeof(dword),[]);
+                      reg_index:=cg.getintregister(current_asmdata.CurrAsmList,OS_INT);
+                      hlcg.a_load_ref_reg(current_asmdata.CurrAsmList,u32inttype,u64inttype,
+                        href_index,reg_index);                                                             // movl _tls_index(%rip), %edx
+                      href.index:=reg_index; href.scalefactor:=8;
+                    end;
+                  current_asmdata.CurrAsmList.concat(taicpu.op_ref_reg(A_MOV,S_Q,href,hregister));         // movq (%rcx), %rcx  |  movq (%rcx,%rdx,8), %rcx
+
+                  reg_offset:=cg.getintregister(current_asmdata.CurrAsmList,OS_INT);
+                  href:=location.reference; // from inherited
+
+                if not (cs_asm_pre_binutils_2_25 in current_settings.globalswitches) then
+                  begin
+                    href.refaddr:=addr_secrel;
+                    current_asmdata.CurrAsmList.concat(taicpu.op_ref_reg(A_MOV,S_L,href,
+                      newreg(getregtype(reg_offset),getsupreg(reg_offset),R_SUBD)));                       // movl $threadvar@secrel32, %rax
+                  end
+                else // Old binutils (2.21) not support $threadvar@secrel32
+                  begin
+                    href.symbol.typ:=AT_DATA;
+                    href.refaddr:=addr_full;
+                    current_asmdata.CurrAsmList.concat(taicpu.op_ref_reg(A_MOV,S_Q,href,reg_offset));      // movq $threadvar, %rax
+                    reg_start:=cg.getintregister(current_asmdata.CurrAsmList,OS_INT);
+                    reference_reset_symbol(href,current_asmdata.RefAsmSymbol('___tls_start__',AT_DATA),0,sizeof(AInt),[]);
+                    href.refaddr:=addr_full;
+                    current_asmdata.CurrAsmList.concat(taicpu.op_ref_reg(A_MOV,S_Q,href,reg_start));       // movq $___tls_start__, %rdx
+                    current_asmdata.CurrAsmList.concat(taicpu.op_reg_reg(A_SUB,S_Q,reg_start,reg_offset)); // subq %rdx, %rax
+                  end;
+
+                  reference_reset(location.reference,location.reference.alignment,location.reference.volatility);
+                  location.reference.base:=hregister;
+                  location.reference.index:=reg_offset;                                                    // (%rcx,%rax)
+                end;
               else
                 ;
             end;
diff --git a/rtl/inc/heaptrc.pp b/rtl/inc/heaptrc.pp
index 920f448293..da7caade9f 100644
--- a/rtl/inc/heaptrc.pp
+++ b/rtl/inc/heaptrc.pp
@@ -1035,11 +1035,13 @@ var
    edata : ptruint; external name '__data_end__';
    sbss : ptruint; external name '__bss_start__';
    ebss : ptruint; external name '__bss_end__';
+{$ifndef FPC_SECTION_THREADVARS}
    TLSKey : PDWord; external name '_FPC_TlsKey';
    TLSSize : DWord; external name '_FPC_TlsSize';
 
 function TlsGetValue(dwTlsIndex : DWord) : pointer;
   {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'TlsGetValue';
+{$endif !FPC_SECTION_THREADVARS}
 {$endif}
 
 {$ifdef BEOS}
@@ -1104,6 +1106,7 @@ begin
   { inside data, rdata ... bss }
   if (ptruint(p)>=ptruint(@sdata)) and (ptruint(p)<ptruint(@ebss)) then
     exit;
+  {$ifndef FPC_SECTION_THREADVARS}
   { is program multi-threaded and p inside Threadvar range? }
   if TlsKey^<>-1 then
     begin
@@ -1112,6 +1115,7 @@ begin
           (ptruint(p)<ptruint(datap)+TlsSize)) then
         exit;
     end;
+  {$endif !FPC_SECTION_THREADVARS}
 {$endif windows}
 
 {$IFDEF OS2}
diff --git a/rtl/linux/si_impl.inc b/rtl/linux/si_impl.inc
index a727310551..e3ceee427e 100644
--- a/rtl/linux/si_impl.inc
+++ b/rtl/linux/si_impl.inc
@@ -20,7 +20,9 @@ procedure SysEntry_InitTLS(constref info: TEntryInformation); external name 'FPC
 
 var
   InitFinalTable : record end; external name 'INITFINAL';
+  {$ifndef FPC_SECTION_THREADVARS}
   ThreadvarTablesTable : record end; external name 'FPC_THREADVARTABLES';
+  {$endif}
   {$ifdef FPC_HAS_RESSTRINITS}
   ResStrInitTables : record end; external name 'FPC_RESSTRINITTABLES';
   {$endif}
diff --git a/rtl/win/sysosh.inc b/rtl/win/sysosh.inc
index 5bee0d4c17..4b65a1d41d 100644
--- a/rtl/win/sysosh.inc
+++ b/rtl/win/sysosh.inc
@@ -56,7 +56,9 @@ type
     {$ifdef WIN32}
     asm_exit : Procedure;stdcall;
     {$endif WIN32}
+    {$ifndef FPC_SECTION_THREADVARS}
     TlsKeyAddr : PDWord;
+    {$endif !FPC_SECTION_THREADVARS}
     SysInstance: {$ifdef CPU64}PQWord{$else}PLongInt{$endif};
     WideInitTables : Pointer;
   end;
@@ -117,7 +119,4 @@ var
 {$ifdef FPC_USE_WIN64_SEH}
 procedure _fpc_local_unwind(frame,target: Pointer);compilerproc;
 {$endif FPC_USE_WIN64_SEH}
-{$ifdef FPC_SECTION_THREADVARS}
-function fpc_tls_add(addr: pointer):pointer;compilerproc;
-{$endif FPC_SECTION_THREADVARS}
 
diff --git a/rtl/win/systhrd.inc b/rtl/win/systhrd.inc
index 10bb07fda4..dfa583fee7 100644
--- a/rtl/win/systhrd.inc
+++ b/rtl/win/systhrd.inc
@@ -19,6 +19,7 @@
                            Local WINApi imports
 *****************************************************************************}
 
+{$ifndef FPC_SECTION_THREADVARS}
 const
   { LocalAlloc flags  }
   LMEM_FIXED = 0;
@@ -34,15 +35,16 @@ function TlsGetValue(dwTlsIndex : DWord) : pointer;
   {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'TlsGetValue';
 function TlsSetValue(dwTlsIndex : DWord;lpTlsValue : pointer) : LongBool;
   {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'TlsSetValue';
+function LocalAlloc(uFlags:DWord; dwBytes:SIZE_T):Pointer;
+  {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'LocalAlloc';
+function LocalFree(hMem : Pointer):Pointer; {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'LocalFree';
+{$endif !FPC_SECTION_THREADVARS}
 function CreateThread(lpThreadAttributes : pointer;
   dwStackSize : SIZE_T; lpStartAddress : pointer;lpParameter : pointer;
   dwCreationFlags : DWord;var lpThreadId : DWord) : THandle;
   {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'CreateThread';
 procedure ExitThread(dwExitCode : DWord);
   {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'ExitThread';
-function LocalAlloc(uFlags:DWord; dwBytes:SIZE_T):Pointer;
-  {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'LocalAlloc';
-function LocalFree(hMem : Pointer):Pointer; {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'LocalFree';
 procedure Sleep(dwMilliseconds: DWord); {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'Sleep';
 function  WinSuspendThread (threadHandle : THandle) : dword; {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'SuspendThread';
 function  WinResumeThread  (threadHandle : THandle) : dword; {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'ResumeThread';
@@ -98,6 +100,10 @@ var
                              Threadvar support
 *****************************************************************************}
 
+    var
+      MainThreadIdWin32 : DWORD;
+      
+{$ifndef FPC_SECTION_THREADVARS}
     var
       // public names are used by heaptrc unit
       threadvarblocksize : dword; public name '_FPC_TlsSize';
@@ -108,9 +114,6 @@ var
       TLSKey : PDWord = @TLSKeyVar; public name '_FPC_TlsKey';
       {$endif FPC_HAS_INDIRECT_ENTRY_INFORMATION}
 
-    var
-      MainThreadIdWin32 : DWORD;
-
     procedure SysInitThreadvar(var offset : dword;size : dword);
       begin
         offset:=threadvarblocksize;
@@ -199,6 +202,7 @@ var
             TlsSetValue(tlskey^, nil);
           end;
       end;
+{$endif !FPC_SECTION_THREADVARS}
 
 
 {*****************************************************************************
@@ -255,7 +259,9 @@ var
         writeln('Creating new thread');
 {$endif DEBUG_MT}
         { Initialize multithreading if not done }
+{$ifndef FPC_SECTION_THREADVARS}
         SysInitTLS;
+{$endif !FPC_SECTION_THREADVARS}
         IsMultiThread:=true;
 
         { the only way to pass data to the newly created thread
@@ -623,10 +629,17 @@ begin
     EnterCriticalSection   :=@SysEnterCriticalSection;
     TryEnterCriticalSection:=@SysTryEnterCriticalSection;
     LeaveCriticalSection   :=@SysLeaveCriticalSection;
+{$ifndef FPC_SECTION_THREADVARS}
     InitThreadVar          :=@SysInitThreadVar;
     RelocateThreadVar      :=@SysRelocateThreadVar;
     AllocateThreadVars     :=@SysAllocateThreadVars;
     ReleaseThreadVars      :=@SysReleaseThreadVars;
+{$else}
+    InitThreadVar          :=Nil;
+    RelocateThreadVar      :=Nil;
+    AllocateThreadVars     :=Nil;
+    ReleaseThreadVars      :=Nil;
+{$endif FPC_SECTION_THREADVARS}
     BasicEventCreate       :=@intBasicEventCreate;
     BasicEventDestroy      :=@intBasicEventDestroy;
     BasicEventResetEvent   :=@intBasicEventResetEvent;
@@ -641,10 +654,12 @@ begin
     end;
   SetThreadManager(WinThreadManager);
   ThreadID := GetCurrentThreadID;
+{$ifndef FPC_SECTION_THREADVARS}
 {$ifndef FPC_USE_TLS_DIRECTORY}
   if IsLibrary then
 {$endif}
     SysInitTLS;
+{$endif !FPC_SECTION_THREADVARS}
 
 {$ifndef WINCE}
   KernelHandle:=GetModuleHandle(KernelDLL);
diff --git a/rtl/win/systlsdir.inc b/rtl/win/systlsdir.inc
index db53b77a94..28bd4e9be3 100644
--- a/rtl/win/systlsdir.inc
+++ b/rtl/win/systlsdir.inc
@@ -32,8 +32,10 @@ Const
   DLL_PROCESS_DETACH = 0;
   DLL_THREAD_DETACH = 3;
 
+{$ifndef FPC_SECTION_THREADVARS}
 var
    TlsKey : PDWord = @TlsKeyVar;
+{$endif !FPC_SECTION_THREADVARS}
 
 type
   TTlsDirectory=packed record
@@ -44,12 +46,14 @@ type
   end;
 
 
+{$ifndef FPC_SECTION_THREADVARS}
 function TlsGetValue(dwTlsIndex : DWord) : pointer; stdcall;
   external 'kernel32' name 'TlsGetValue';
 
 procedure InitSystemThreads; external name '_FPC_InitSystemThreads';
 procedure SysAllocateThreadVars; external name '_FPC_SysAllocateThreadVars';
 procedure InitHeap; external name '_FPC_InitHeap';
+{$endif !FPC_SECTION_THREADVARS}
 
 {$ifdef FPC_HAS_INDIRECT_ENTRY_INFORMATION}
 procedure SetupEntryInformation(constref info: TEntryInformation); external name '_FPC_SetupEntryInformation';
@@ -58,7 +62,7 @@ procedure SetupEntryInformation(constref info: TEntryInformation); external name
 {$endif FPC_INSSIDE_SYSINIT}
 
 procedure Exec_Tls_callback(Handle : pointer; reason : Dword; Reserved : pointer);
-  stdcall; [public,alias:'_FPC_Tls_Callback'];
+  stdcall;
   begin
      if IsLibrary then
        Exit;
@@ -77,6 +81,7 @@ procedure Exec_Tls_callback(Handle : pointer; reason : Dword; Reserved : pointer
          and the EntryInformation is a constant which sholud prevent troubles }
        DLL_PROCESS_ATTACH:
          begin
+         {$ifndef FPC_SECTION_THREADVARS}
            {$ifdef FPC_HAS_INDIRECT_ENTRY_INFORMATION}
            { since this procedure is called before SetupEntryInformation and thus
              before EXE_Entry we need to setup the entry information here so that
@@ -85,10 +90,12 @@ procedure Exec_Tls_callback(Handle : pointer; reason : Dword; Reserved : pointer
            {$endif FPC_HAS_INDIRECT_ENTRY_INFORMATION}
            InitHeap;
            InitSystemThreads;
+         {$endif !FPC_SECTION_THREADVARS}
          end;
 
        DLL_THREAD_ATTACH :
          begin
+         {$ifndef FPC_SECTION_THREADVARS}
          {  !!! SysInitMultithreading must NOT be called here. Windows guarantees that
             the main thread invokes PROCESS_ATTACH, not THREAD_ATTACH. So this always
             executes in non-main thread. SysInitMultithreading() here will cause
@@ -100,6 +107,7 @@ procedure Exec_Tls_callback(Handle : pointer; reason : Dword; Reserved : pointer
          }
            { Allocate Threadvars  }
            SysAllocateThreadVars;
+         {$endif !FPC_SECTION_THREADVARS}
 
            { NS : no idea what is correct to pass here - pass dummy value for now }
            { passing a dummy is ok, the correct value is read from the coff header of SysInstance (FK) }
@@ -107,7 +115,11 @@ procedure Exec_Tls_callback(Handle : pointer; reason : Dword; Reserved : pointer
          end;
        DLL_THREAD_DETACH :
          begin
+         {$ifdef FPC_SECTION_THREADVARS}
+           if ThreadID<>0 then
+         {$else !FPC_SECTION_THREADVARS}
            if TlsGetValue(TLSKey^)<>nil then
+         {$endif !FPC_SECTION_THREADVARS}
              DoneThread; { Assume everything is idempotent there }
          end;
      end;
diff --git a/rtl/win/syswin.inc b/rtl/win/syswin.inc
index 63ee752569..37fb574f33 100644
--- a/rtl/win/syswin.inc
+++ b/rtl/win/syswin.inc
@@ -408,10 +408,12 @@ function Dll_entry{$ifdef FPC_HAS_INDIRECT_ENTRY_INFORMATION}(constref info : TE
          end;
        DLL_THREAD_ATTACH :
          begin
+         {$ifndef FPC_SECTION_THREADVARS}
            { SysInitMultithreading must not be called here,
              see comments in exec_tls_callback below }
            { Allocate Threadvars  }
            SysAllocateThreadVars;
+         {$endif !FPC_SECTION_THREADVARS}
 
            { NS : no idea what is correct to pass here - pass dummy value for now }
            { passing a dummy is ok, the correct value is read from the coff header of SysInstance (FK) }
@@ -425,7 +427,11 @@ function Dll_entry{$ifdef FPC_HAS_INDIRECT_ENTRY_INFORMATION}(constref info : TE
            if assigned(Dll_Thread_Detach_Hook) then
              Dll_Thread_Detach_Hook(DllParam);
            { Release Threadvars }
+         {$ifdef FPC_SECTION_THREADVARS}
+           if ThreadID<>0 then
+         {$else !FPC_SECTION_THREADVARS}
            if TlsGetValue(TLSKey^)<>nil then
+         {$endif !FPC_SECTION_THREADVARS}
              DoneThread; { Assume everything is idempotent there }
          end;
        DLL_PROCESS_DETACH :
@@ -439,9 +445,11 @@ function Dll_entry{$ifdef FPC_HAS_INDIRECT_ENTRY_INFORMATION}(constref info : TE
                InternalExit;
              end;
 
+         {$ifndef FPC_SECTION_THREADVARS}
            SysReleaseThreadVars;
            { Free TLS resources used by ThreadVars }
            SysFiniMultiThreading;
+         {$endif !FPC_SECTION_THREADVARS}
            MainThreadIDWin32:=0;
          end;
      end;
diff --git a/rtl/win32/sysinit.inc b/rtl/win32/sysinit.inc
index 0eea501af0..c8f7b6f042 100644
--- a/rtl/win32/sysinit.inc
+++ b/rtl/win32/sysinit.inc
@@ -16,10 +16,14 @@
 
    var
       SysInstance : LongInt;
+ {$ifndef FPC_SECTION_THREADVARS}
       TlsKeyVar: DWord = $ffffffff;
+ {$endif !FPC_SECTION_THREADVARS}
 
       InitFinalTable : record end; external name 'INITFINAL';
+ {$ifndef FPC_SECTION_THREADVARS}
       ThreadvarTablesTable : record end; external name 'FPC_THREADVARTABLES';
+ {$endif !FPC_SECTION_THREADVARS}
       WideInitTables : record end; external name 'FPC_WIDEINITTABLES';
  {$ifdef FPC_HAS_RESSTRINITS}
       ResStrInitTables : record end; external name 'FPC_RESSTRINITTABLES';
@@ -53,7 +57,11 @@
       STD_INPUT_HANDLE = dword(-10);
       SysInitEntryInformation : TEntryInformation = (
         InitFinalTable : @InitFinalTable;
+{$ifdef FPC_SECTION_THREADVARS}
+        ThreadvarTablesTable : nil;
+{$else FPC_SECTION_THREADVARS}
         ThreadvarTablesTable : @ThreadvarTablesTable;
+{$endif FPC_SECTION_THREADVARS}
         ResourceStringTables : @ResourceStringTables;
 {$ifdef FPC_HAS_RESSTRINITS}
         ResStrInitTables : @ResStrInitTables;
@@ -65,7 +73,9 @@
         valgrind_used : false;
         OS : (
           asm_exit : @asm_exit;
+{$ifndef FPC_SECTION_THREADVARS}
           TlsKeyAddr : @TlsKeyVar;
+{$endif !FPC_SECTION_THREADVARS}
           SysInstance : @SysInstance;
           WideInitTables: @WideInitTables;
           );
diff --git a/rtl/win32/system.pp b/rtl/win32/system.pp
index 4ffa0b65ff..4cf69de94d 100644
--- a/rtl/win32/system.pp
+++ b/rtl/win32/system.pp
@@ -78,7 +78,9 @@ end;
 
 procedure OsSetupEntryInformation(constref info: TEntryInformation);
 begin
+{$ifndef FPC_SECTION_THREADVARS}
   TlsKey := info.OS.TlsKeyAddr;
+{$endif !FPC_SECTION_THREADVARS}
   FPCSysInstance := info.OS.SysInstance;
   WStrInitTablesTable := info.OS.WideInitTables;
 end;
@@ -498,41 +500,6 @@ begin
 end;
 {$endif Set_i386_Exception_handler}
 
-{$ifdef FPC_SECTION_THREADVARS}
-function fpc_tls_add(addr: pointer): pointer; assembler; nostackframe;
-  [public,alias: 'FPC_TLS_ADD']; compilerproc;
-  asm
-      sub   $tls_data_start,%eax
-      cmpb  $0,IsLibrary
-      mov   _tls_index,%ecx
-      jnz   .L1
-      mov   %fs:(0x2c),%edx
-      add   (%edx,%ecx,4),%eax
-      ret
-.L1:
-      push  %ebx
-      mov   %eax,%ebx
-      call  GetLastError
-      push  %eax                      { save LastError }
-      push  _tls_index
-      call  TlsGetValue
-      test  %eax,%eax
-      jnz   .L2
-      { This can happen when a thread existed before DLL was loaded,
-        or if DisableThreadLibraryCalls was called. }
-      call  SysAllocateThreadVars
-      mov   $0x1000000,%eax
-      call  InitThread
-      push  _tls_index
-      call  TlsGetValue
-.L2:
-      add   %eax,%ebx
-      call  SetLastError              { restore (value is on stack) }
-      mov   %ebx,%eax
-      pop   %ebx
-  end;
-{$endif FPC_SECTION_THREADVARS}
-
 function CheckInitialStkLen(stklen : SizeUInt) : SizeUInt;
   type
     tdosheader = packed record
diff --git a/rtl/win64/sysinit.pp b/rtl/win64/sysinit.pp
index 5d396beb8f..9d740e19ec 100644
--- a/rtl/win64/sysinit.pp
+++ b/rtl/win64/sysinit.pp
@@ -21,10 +21,14 @@ unit sysinit;
 
    var
       SysInstance : QWord;
+{$ifndef FPC_SECTION_THREADVARS}
       TlsKeyVar: DWord = $ffffffff;
+{$endif !FPC_SECTION_THREADVARS}
 
       InitFinalTable : record end; external name 'INITFINAL';
+{$ifndef FPC_SECTION_THREADVARS}
       ThreadvarTablesTable : record end; external name 'FPC_THREADVARTABLES';
+{$endif !FPC_SECTION_THREADVARS}
       WideInitTables : record end; external name 'FPC_WIDEINITTABLES';
 {$ifdef FPC_HAS_RESSTRINITS}
       ResStrInitTables : record end; external name 'FPC_RESSTRINITTABLES';
@@ -58,7 +62,11 @@ unit sysinit;
       STD_INPUT_HANDLE = dword(-10);
       SysInitEntryInformation : TEntryInformation = (
         InitFinalTable : @InitFinalTable;
+{$ifdef FPC_SECTION_THREADVARS}
+        ThreadvarTablesTable : nil;
+{$else FPC_SECTION_THREADVARS}
         ThreadvarTablesTable : @ThreadvarTablesTable;
+{$endif FPC_SECTION_THREADVARS}
         ResourceStringTables : @ResourceStringTables;
 {$ifdef FPC_HAS_RESSTRINITS}
         ResStrInitTables : @ResStrInitTables;
@@ -69,7 +77,9 @@ unit sysinit;
         PascalMain : @PascalMain;
         valgrind_used : false;
         OS : (
+{$ifndef FPC_SECTION_THREADVARS}
           TlsKeyAddr : @TlsKeyVar;
+{$endif !FPC_SECTION_THREADVARS}
           SysInstance : @SysInstance;
           WideInitTables: @WideInitTables;
           );
@@ -124,6 +134,9 @@ unit sysinit;
       sysinstance:=_hinstance;
       dllreason:=_dllreason;
       dllparam:=PtrInt(_dllparam);
+{$if defined(FPC_USE_TLS_DIRECTORY) and defined(FPC_SECTION_THREADVARS)}
+      LinkIn(@tlsdir,@tls_callback_end,@tls_callback);
+{$endif}
       SetupEntryInformation;
       DLL_Entry(SysInitEntryInformation);
     end;
@@ -135,6 +148,9 @@ unit sysinit;
       sysinstance:=_hinstance;
       dllreason:=_dllreason;
       dllparam:=PtrInt(_dllparam);
+{$if defined(FPC_USE_TLS_DIRECTORY) and defined(FPC_SECTION_THREADVARS)}
+      LinkIn(@tlsdir,@tls_callback_end,@tls_callback);
+{$endif}
       SetupEntryInformation;
       DLL_Entry(SysInitEntryInformation);
     end;
diff --git a/rtl/win64/system.pp b/rtl/win64/system.pp
index 6cf7b174cd..b6fad06fff 100644
--- a/rtl/win64/system.pp
+++ b/rtl/win64/system.pp
@@ -122,7 +122,9 @@ procedure PascalMain;external name 'PASCALMAIN';
 
 procedure OsSetupEntryInformation(constref info: TEntryInformation);
 begin
+{$ifndef FPC_SECTION_THREADVARS}
   TlsKey := info.OS.TlsKeyAddr;
+{$endif !FPC_SECTION_THREADVARS}
   FPCSysInstance := info.OS.SysInstance;
   WStrInitTablesTable := info.OS.WideInitTables;
 end;
@@ -504,47 +506,6 @@ begin
 end;
 {$endif VER3_0}
 
-{$ifdef FPC_SECTION_THREADVARS}
-function fpc_tls_add(addr: pointer): pointer; assembler; nostackframe;
-  [public,alias: 'FPC_TLS_ADD']; compilerproc;
-  asm
-      sub   $56,%rsp                  { 32 spill area + 16 local vars + 8 misalignment }
-  .seh_stackalloc 56
-  .seh_endprologue
-      lea   tls_data_start(%rip),%rax
-      sub   %rax,%rcx
-      cmpb  $0,IsLibrary(%rip)
-      mov   _tls_index(%rip),%eax
-      jnz   .L1
-      mov   %gs:(88),%rdx
-      add   (%rdx,%rax,8),%rcx
-      mov   %rcx,%rax
-      jmp   .L3
-.L1:
-      mov   %rcx,32(%rsp)
-      call  GetLastError
-      mov   %rax,40(%rsp)             { save LastError }
-      mov   _tls_index(%rip),%ecx
-      call  TlsGetValue
-      test  %rax,%rax
-      jnz   .L2
-      { This can happen when a thread existed before DLL was loaded,
-        or if DisableThreadLibraryCalls was called. }
-      call  SysAllocateThreadVars
-      mov   $0x1000000,%rcx
-      call  InitThread
-      mov   _tls_index(%rip),%ecx
-      call  TlsGetValue
-.L2:
-      add   %rax,32(%rsp)
-      mov   40(%rsp),%rcx
-      call  SetLastError
-      mov   32(%rsp),%rax
-.L3:
-      add   $56,%rsp
-  end;
-{$endif FPC_SECTION_THREADVARS}
-
 function CheckInitialStkLen(stklen : SizeUInt) : SizeUInt;
   type
     tdosheader = packed record
-- 
2.26.2.windows.1

package Rtl;

contains
  System,
  ObjPas,
  MacPas,
  Iso7185,
  FPIntRes,
{$ifdef darwin}
  FPExtRes,
{$endif}
  Classes,
  SysUtils,
  character,
  charset,
  cpu,
  dos,
  dynlibs,
  exeinfo,
  fgl,
  fpwidestring,
  getopts,
  heaptrc,
  lineinfo,
  lnfodwrf,
  math,
  rtlconsts,
  strings,
  sysconst,
  types,
  typinfo,
  uuchar,
{$ifdef linux}
  syscall,
  linux,
  linuxvcs,
{$endif}
{$ifdef unix}
  ports,
  errors,
  dl,
  unixcp,
  unix,
  unixtype,
  unixutil,
  termio,
  baseunix,
  x86,
{$endif}
{$ifdef cpux86}
  mmx,
{$endif}
{$ifdef windows}
  Windows,
{$endif}
  { some unit at the end after the defines }
  ctypes;

end.

---
 compiler/ogcoff.pas        |  7 ++++++-
 compiler/systems/t_win.pas | 27 ++++++++++++++++++++++++++-
 rtl/win/systlsdir.inc      |  8 ++++++--
 3 files changed, 38 insertions(+), 4 deletions(-)

diff --git a/compiler/ogcoff.pas b/compiler/ogcoff.pas
index 803b9ebad3..90bbca8b0d 100644
--- a/compiler/ogcoff.pas
+++ b/compiler/ogcoff.pas
@@ -2776,7 +2776,10 @@ const pemagic : array[0..3] of byte = (
               end
             else
               begin
-                peoptheader.MajorOperatingSystemVersion:=4;
+                if tf_section_threadvars in target_info.flags then
+                  peoptheader.MajorOperatingSystemVersion:=6 // >= Windows Vista
+                else
+                  peoptheader.MajorOperatingSystemVersion:=4;
                 peoptheader.MinorOperatingSystemVersion:=0;
               end;
             if SetPEUserVersionSetExplicitely then
@@ -2798,6 +2801,8 @@ const pemagic : array[0..3] of byte = (
               begin
                 if target_info.system in systems_wince then
                   peoptheader.MajorSubsystemVersion:=3
+                else if tf_section_threadvars in target_info.flags then
+                  peoptheader.MajorSubsystemVersion:=6 // >= Windows Vista
                 else
                   peoptheader.MajorSubsystemVersion:=4;
                 peoptheader.MinorSubsystemVersion:=0;
diff --git a/compiler/systems/t_win.pas b/compiler/systems/t_win.pas
index 887787f8e9..b96f820a3c 100644
--- a/compiler/systems/t_win.pas
+++ b/compiler/systems/t_win.pas
@@ -1711,11 +1711,36 @@ implementation
             else
               ;
           end;
-        if dllversion<>'' then
+        if SetPEUserVersionSetExplicitely then
+          begin
+            peoptheader.MajorImageVersion:=peuserversionmajor;
+            peoptheader.MinorImageVersion:=peuserversionminor;
+          end
+        else if dllversion<>'' then
           begin
            peoptheader.MajorImageVersion:=dllmajor;
            peoptheader.MinorImageVersion:=dllminor;
           end;
+        if SetPEOSVersionSetExplicitely then
+          begin
+            peoptheader.MajorOperatingSystemVersion:=peosversionmajor;
+            peoptheader.MinorOperatingSystemVersion:=peosversionminor;
+          end
+        else if tf_section_threadvars in target_info.flags then
+          begin
+            peoptheader.MajorOperatingSystemVersion:=6; // >= Windows Vista
+            peoptheader.MinorOperatingSystemVersion:=0;
+          end;
+        if SetPESubSysVersionSetExplicitely then
+          begin
+            peoptheader.MajorSubsystemVersion:=pesubsysversionmajor;
+            peoptheader.MinorSubsystemVersion:=pesubsysversionminor;
+          end
+        else if tf_section_threadvars in target_info.flags then
+          begin
+            peoptheader.MajorSubsystemVersion:=6; // >= Windows Vista
+            peoptheader.MinorSubsystemVersion:=0;
+          end;
         { reset timestamp }
         peheader.time:=0;
         { write header back, skip pe magic }
diff --git a/rtl/win/systlsdir.inc b/rtl/win/systlsdir.inc
index 28bd4e9be3..d0b48a89a5 100644
--- a/rtl/win/systlsdir.inc
+++ b/rtl/win/systlsdir.inc
@@ -52,8 +52,8 @@ function TlsGetValue(dwTlsIndex : DWord) : pointer; stdcall;
 
 procedure InitSystemThreads; external name '_FPC_InitSystemThreads';
 procedure SysAllocateThreadVars; external name '_FPC_SysAllocateThreadVars';
-procedure InitHeap; external name '_FPC_InitHeap';
 {$endif !FPC_SECTION_THREADVARS}
+procedure InitHeap; external name '_FPC_InitHeap';
 
 {$ifdef FPC_HAS_INDIRECT_ENTRY_INFORMATION}
 procedure SetupEntryInformation(constref info: TEntryInformation); external name '_FPC_SetupEntryInformation';
@@ -81,14 +81,18 @@ procedure Exec_Tls_callback(Handle : pointer; reason : Dword; Reserved : pointer
          and the EntryInformation is a constant which sholud prevent troubles }
        DLL_PROCESS_ATTACH:
          begin
-         {$ifndef FPC_SECTION_THREADVARS}
            {$ifdef FPC_HAS_INDIRECT_ENTRY_INFORMATION}
            { since this procedure is called before SetupEntryInformation and thus
              before EXE_Entry we need to setup the entry information here so that
              the threadvar handling can be correctly initialized }
            SetupEntryInformation(SysInitEntryInformation);
            {$endif FPC_HAS_INDIRECT_ENTRY_INFORMATION}
+           { the DLL_PROCESS_ATTACH initialization for the heap is required, because
+             if the DLL_PROCESS_ATTACH entry of a used DLL creates a thread, DLL_THREAD_ATTACH
+             of the executable will be called so that the RTL for the thread can be initialized
+             and that needs the code in DLL_PROCESS_ATTACH to be run first }
            InitHeap;
+         {$ifndef FPC_SECTION_THREADVARS}
            InitSystemThreads;
          {$endif !FPC_SECTION_THREADVARS}
          end;
-- 
2.26.2.windows.1

---
 compiler/aggas.pas         |  4 +-
 compiler/cgbase.pas        |  1 +
 compiler/dbgdwarf.pas      |  3 +-
 compiler/nld.pas           |  4 +-
 compiler/nutils.pas        |  9 +++--
 compiler/ogcoff.pas        | 35 +++++++----------
 compiler/systems/i_win.pas |  3 ++
 compiler/systems/t_win.pas | 27 ++++++++++++-
 compiler/x86/aasmcpu.pas   |  6 +++
 compiler/x86/agx86att.pas  |  9 +++++
 compiler/x86/nx86ld.pas    | 79 +++++++++++++++++++++++---------------
 rtl/inc/heaptrc.pp         |  4 ++
 rtl/linux/si_impl.inc      |  2 +
 rtl/win/sysosh.inc         |  5 +--
 rtl/win/systhrd.inc        | 27 ++++++++++---
 rtl/win/systlsdir.inc      | 18 ++++++++-
 rtl/win/syswin.inc         |  8 ++++
 rtl/win32/sysinit.inc      | 10 +++++
 rtl/win32/system.pp        | 37 +-----------------
 rtl/win64/sysinit.pp       | 16 ++++++++
 rtl/win64/system.pp        | 43 +--------------------
 21 files changed, 203 insertions(+), 147 deletions(-)

diff --git a/compiler/aggas.pas b/compiler/aggas.pas
index 41d20c5129..f9b69f605b 100644
--- a/compiler/aggas.pas
+++ b/compiler/aggas.pas
@@ -533,7 +533,9 @@ implementation
              { according to the GNU AS guide AS for COFF does not support the
                progbits }
              writer.AsmWrite('.section ');
-             usesectionflags:=true;
+             { .tls section requires special processing }
+             if not((atype=sec_threadvar) and (tf_section_threadvars in target_info.flags)) then
+               usesectionflags:=true;
            end;
          system_powerpc_darwin,
          system_i386_darwin,
diff --git a/compiler/cgbase.pas b/compiler/cgbase.pas
index 6dca71ba40..0e862ca739 100644
--- a/compiler/cgbase.pas
+++ b/compiler/cgbase.pas
@@ -140,6 +140,7 @@ interface
 {$ifdef x86_64}
           ,addr_tpoff
           ,addr_tlsgd
+          ,addr_secrel
 {$endif x86_64}
          );
 
diff --git a/compiler/dbgdwarf.pas b/compiler/dbgdwarf.pas
index 5b75e257b5..d81436ed2a 100644
--- a/compiler/dbgdwarf.pas
+++ b/compiler/dbgdwarf.pas
@@ -2658,7 +2658,8 @@ implementation
                   begin
                     if vo_is_thread_var in sym.varoptions then
                       begin
-                        if tf_section_threadvars in target_info.flags then
+                        if (tf_section_threadvars in target_info.flags) and
+                           (not (target_info.system in [system_x86_64_win64])) then // TODO : dwarf for threadvars on win64
                           begin
                             case sizeof(puint) of
                               2:
diff --git a/compiler/nld.pas b/compiler/nld.pas
index e2bccbc080..0453f162e8 100644
--- a/compiler/nld.pas
+++ b/compiler/nld.pas
@@ -450,7 +450,9 @@ implementation
                 { call to get address of threadvar }
                 if (vo_is_thread_var in tabstractvarsym(symtableentry).varoptions) then
                   begin
-                    include(current_procinfo.flags,pi_do_call);
+                    if (not (tf_section_threadvars in target_info.flags)) or
+                       (current_settings.tlsmodel in [tlsm_local_dynamic,tlsm_global_dynamic]) then
+                      include(current_procinfo.flags,pi_do_call);
                     include(current_procinfo.flags,pi_uses_threadvar);
                   end;
               end;
diff --git a/compiler/nutils.pas b/compiler/nutils.pas
index d586fc222d..b5c0de1937 100644
--- a/compiler/nutils.pas
+++ b/compiler/nutils.pas
@@ -191,7 +191,7 @@ interface
 implementation
 
     uses
-      cutils,verbose,globals,
+      cutils,verbose,globals,systems,
       symconst,symdef,
       defcmp,defutil,
       ncon,ncnv,nld,nflw,nset,ncal,nadd,nmem,ninl,
@@ -799,10 +799,13 @@ implementation
                 begin
                   if assigned(tloadnode(p).left) then
                     inc(result,node_complexity(tloadnode(p).left));
-                  { threadvars need a helper call }
+                  { some threadvars models need a helper call }
                   if (tloadnode(p).symtableentry.typ=staticvarsym) and
                      (vo_is_thread_var in tstaticvarsym(tloadnode(p).symtableentry).varoptions) then
-                    inc(result,5)
+                    if (tf_section_threadvars in target_info.flags) and (current_settings.tlsmodel in [tlsm_local_exec,tlsm_initial_exec]) then
+                      inc(result,3)
+                    else
+                      inc(result,50)
                   else if not((tloadnode(p).symtableentry.typ in [staticvarsym,localvarsym,paravarsym,fieldvarsym]) and
                     (tabstractvarsym(tloadnode(p).symtableentry).varregable in [vr_intreg,vr_mmreg,vr_fpureg])) then
                     inc(result);
diff --git a/compiler/ogcoff.pas b/compiler/ogcoff.pas
index 8c05b3792a..90bbca8b0d 100644
--- a/compiler/ogcoff.pas
+++ b/compiler/ogcoff.pas
@@ -1015,8 +1015,10 @@ const pemagic : array[0..3] of byte = (
                   end;
                 RELOC_SECREL32 :
                   begin
-                    { fixup address when the symbol was known in defined object }
-                    if (relocsec.objdata=objsec.objdata) then
+                    { fixup address when the symbol was known in defined object or
+                      when the symbol is a threadvar }
+                    if (relocsec.objdata=objsec.objdata) or
+                       assigned(objreloc.symbol) and (copy(objreloc.symbol.objsection.Name,1,4)='.tls') then
                       dec(address,relocsec.ExeSection.MemPos);
                     inc(address,relocval);
                   end;
@@ -2667,21 +2669,17 @@ const pemagic : array[0..3] of byte = (
 
         procedure UpdateTlsDataDir;
         var
-          {callbacksection : TExeSection;}
           tlsexesymbol: TExeSymbol;
           tlssymbol: TObjSymbol;
-          callbackexesymbol: TExeSymbol;
-          //callbacksymbol: TObjSymbol;
         begin
           { according to GNU ld,
             the callback routines should be placed into .CRT$XL*
             sections, and the thread local variables in .tls
             __tls_start__ and __tls_end__ symbols
-            should be used for the initialized part,
-            which we do not support yet. }
-          { For now, we only pass the address of the __tls_used
+            should be used for the initialized part. }
+          { We pass the address of the __tls_used
             asm symbol into PE_DATADIR_TLS with the correct
-            size of this table (different for win32/win64 }
+            size of this table (different for win32/win64) }
           tlsexesymbol:=texesymbol(ExeSymbolList.Find(
             target_info.Cprefix+'_tls_used'));
           if assigned(tlsexesymbol) then
@@ -2690,18 +2688,6 @@ const pemagic : array[0..3] of byte = (
               peoptheader.DataDirectory[PE_DATADIR_TLS].vaddr:=tlssymbol.address;
               { sizeof(TlsDirectory) is different on host and target when cross-compiling }
               peoptheader.DataDirectory[PE_DATADIR_TLS].size:=TLSDIR_SIZE;
-              if IsSharedLibrary then
-                begin
-                  { Here we should reset __FPC_tls_callbacks value to nil }
-                  callbackexesymbol:=texesymbol(ExeSymbolList.Find(
-                                        '__FPC_tls_callbacks'));
-                  if assigned (callbackexesymbol) then
-                    begin
-                      //callbacksymbol:=callbackexesymbol.ObjSymbol;
-
-                    end;
-                end;
-
            end;
         end;
 
@@ -2790,7 +2776,10 @@ const pemagic : array[0..3] of byte = (
               end
             else
               begin
-                peoptheader.MajorOperatingSystemVersion:=4;
+                if tf_section_threadvars in target_info.flags then
+                  peoptheader.MajorOperatingSystemVersion:=6 // >= Windows Vista
+                else
+                  peoptheader.MajorOperatingSystemVersion:=4;
                 peoptheader.MinorOperatingSystemVersion:=0;
               end;
             if SetPEUserVersionSetExplicitely then
@@ -2812,6 +2801,8 @@ const pemagic : array[0..3] of byte = (
               begin
                 if target_info.system in systems_wince then
                   peoptheader.MajorSubsystemVersion:=3
+                else if tf_section_threadvars in target_info.flags then
+                  peoptheader.MajorSubsystemVersion:=6 // >= Windows Vista
                 else
                   peoptheader.MajorSubsystemVersion:=4;
                 peoptheader.MinorSubsystemVersion:=0;
diff --git a/compiler/systems/i_win.pas b/compiler/systems/i_win.pas
index d4e8f842d5..d1ffb4e57c 100644
--- a/compiler/systems/i_win.pas
+++ b/compiler/systems/i_win.pas
@@ -107,6 +107,9 @@ unit i_win;
             shortname    : 'Win64';
             flags        : [tf_files_case_aware,tf_has_dllscanner,
                             tf_smartlink_sections,
+{$if not defined(DISABLE_TLS_DIRECTORY) and not defined(SUPPORT_WINXP)}
+                            tf_section_threadvars,
+{$endif}
                             tf_winlikewidestring,tf_no_pic_supported,
                             tf_dwarf_only_local_labels,
                             tf_no_generic_stackcheck,tf_has_winlike_resources,
diff --git a/compiler/systems/t_win.pas b/compiler/systems/t_win.pas
index 887787f8e9..b96f820a3c 100644
--- a/compiler/systems/t_win.pas
+++ b/compiler/systems/t_win.pas
@@ -1711,11 +1711,36 @@ implementation
             else
               ;
           end;
-        if dllversion<>'' then
+        if SetPEUserVersionSetExplicitely then
+          begin
+            peoptheader.MajorImageVersion:=peuserversionmajor;
+            peoptheader.MinorImageVersion:=peuserversionminor;
+          end
+        else if dllversion<>'' then
           begin
            peoptheader.MajorImageVersion:=dllmajor;
            peoptheader.MinorImageVersion:=dllminor;
           end;
+        if SetPEOSVersionSetExplicitely then
+          begin
+            peoptheader.MajorOperatingSystemVersion:=peosversionmajor;
+            peoptheader.MinorOperatingSystemVersion:=peosversionminor;
+          end
+        else if tf_section_threadvars in target_info.flags then
+          begin
+            peoptheader.MajorOperatingSystemVersion:=6; // >= Windows Vista
+            peoptheader.MinorOperatingSystemVersion:=0;
+          end;
+        if SetPESubSysVersionSetExplicitely then
+          begin
+            peoptheader.MajorSubsystemVersion:=pesubsysversionmajor;
+            peoptheader.MinorSubsystemVersion:=pesubsysversionminor;
+          end
+        else if tf_section_threadvars in target_info.flags then
+          begin
+            peoptheader.MajorSubsystemVersion:=6; // >= Windows Vista
+            peoptheader.MinorSubsystemVersion:=0;
+          end;
         { reset timestamp }
         peheader.time:=0;
         { write header back, skip pe magic }
diff --git a/compiler/x86/aasmcpu.pas b/compiler/x86/aasmcpu.pas
index f712af7196..115474e6d1 100644
--- a/compiler/x86/aasmcpu.pas
+++ b/compiler/x86/aasmcpu.pas
@@ -3617,6 +3617,12 @@ implementation
                       currabsreloc:=RELOC_TLSGD;
                       currabsreloc32:=RELOC_TLSGD;
                     end
+                  else if oper[opidx]^.ref^.refaddr=addr_secrel then
+                    begin
+                      currrelreloc:=RELOC_SECREL32;
+                      currabsreloc:=RELOC_SECREL32;
+                      currabsreloc32:=RELOC_SECREL32;
+                    end
                   else
 {$endif x86_64}
                     begin
diff --git a/compiler/x86/agx86att.pas b/compiler/x86/agx86att.pas
index 657f073163..bf69e08bd5 100644
--- a/compiler/x86/agx86att.pas
+++ b/compiler/x86/agx86att.pas
@@ -190,6 +190,8 @@ interface
                owner.writer.AsmWrite('@tpoff');
              addr_tlsgd:
                owner.writer.AsmWrite('@tlsgd');
+             addr_secrel:
+               owner.writer.AsmWrite('@secrel32');
 {$endif x86_64}
              else
                ;
@@ -259,6 +261,13 @@ interface
               {$ifdef x86_64},addr_tpoff,addr_tlsgd{$endif x86_64}
               ] then
               WriteReference(o.ref^)
+        {$ifdef x86_64}
+            else if o.ref^.refaddr=addr_secrel then
+              begin
+                owner.writer.AsmWrite('$');
+                WriteReference(o.ref^);
+              end
+        {$endif x86_64}
             else
               begin
                 owner.writer.AsmWrite('$');
diff --git a/compiler/x86/nx86ld.pas b/compiler/x86/nx86ld.pas
index 774addd1ca..61c82755d4 100644
--- a/compiler/x86/nx86ld.pas
+++ b/compiler/x86/nx86ld.pas
@@ -38,10 +38,10 @@ interface
 implementation
 
     uses
-      globals,
+      globals,fmodule,
       cutils,verbose,systems,
       aasmbase,aasmtai,aasmdata,aasmcpu,
-      cgutils,cgobj,
+      cgutils,cgobj,hlcgobj,
       symconst,symdef,symtable,
       cgbase,cpubase,parabase,paramgr,
       procinfo;
@@ -54,37 +54,10 @@ implementation
       var
         paraloc1 : tcgpara;
         pd: tprocdef;
-        href: treference;
-        hregister : tregister;
-        handled: boolean;
+        href, href_index: treference;
+        hregister, reg_index, reg_offset, reg_start : tregister;
       begin
-        handled:=false;
-        if (tf_section_threadvars in target_info.flags) then
-          begin
-            if target_info.system in [system_i386_win32,system_x86_64_win64] then
-              begin
-                paraloc1.init;
-                pd:=search_system_proc('fpc_tls_add');
-                paramanager.getcgtempparaloc(current_asmdata.CurrAsmList,pd,1,paraloc1);
-                if not(vo_is_weak_external in gvs.varoptions) then
-                  reference_reset_symbol(href,current_asmdata.RefAsmSymbol(gvs.mangledname,AT_DATA,use_indirect_symbol(gvs)),0,sizeof(pint),[])
-                else
-                  reference_reset_symbol(href,current_asmdata.WeakRefAsmSymbol(gvs.mangledname,AT_DATA),0,sizeof(pint),[]);
-                cg.a_loadaddr_ref_cgpara(current_asmdata.CurrAsmList,href,paraloc1);
-                paramanager.freecgpara(current_asmdata.CurrAsmList,paraloc1);
-                paraloc1.done;
-
-                cg.g_call(current_asmdata.CurrAsmList,'FPC_TLS_ADD');
-                cg.ungetcpuregister(current_asmdata.CurrAsmList,NR_FUNCTION_RESULT_REG);
-                hregister:=cg.getaddressregister(current_asmdata.CurrAsmList);
-                cg.a_load_reg_reg(current_asmdata.CurrAsmList,OS_ADDR,OS_ADDR,NR_FUNCTION_RESULT_REG,hregister);
-                location.reference.base:=hregister;
-                handled:=true;
-              end;
-          end;
-
-        if not handled then
-          inherited;
+        inherited;
 
         if (tf_section_threadvars in target_info.flags) then
           begin
@@ -156,6 +129,48 @@ implementation
                       Internalerror(2019012002);
                   end;
                 end;
+              system_x86_64_win64:
+                begin
+                  reference_reset(href,sizeof(AInt),[]); href.segment:=NR_GS; href.offset:=88;
+                  hregister:=cg.getaddressregister(current_asmdata.CurrAsmList);
+                  current_asmdata.CurrAsmList.concat(taicpu.op_ref_reg(A_MOV,S_Q,href,hregister));         // movq %gs:88, %rcx
+
+                  reference_reset_base(href,hregister,0,ctempposinvalid,sizeof(AInt),[]);
+                  if (main_module=nil) or main_module.IsLibrary or main_module.is_unit then
+                    begin // _tls_index <> 0
+                      reference_reset_symbol(href_index,current_asmdata.RefAsmSymbol('_tls_index',AT_DATA),0,sizeof(dword),[]);
+                      reg_index:=cg.getintregister(current_asmdata.CurrAsmList,OS_INT);
+                      hlcg.a_load_ref_reg(current_asmdata.CurrAsmList,u32inttype,u64inttype,
+                        href_index,reg_index);                                                             // movl _tls_index(%rip), %edx
+                      href.index:=reg_index; href.scalefactor:=8;
+                    end;
+                  current_asmdata.CurrAsmList.concat(taicpu.op_ref_reg(A_MOV,S_Q,href,hregister));         // movq (%rcx), %rcx  |  movq (%rcx,%rdx,8), %rcx
+
+                  reg_offset:=cg.getintregister(current_asmdata.CurrAsmList,OS_INT);
+                  href:=location.reference; // from inherited
+
+                if not (cs_asm_pre_binutils_2_25 in current_settings.globalswitches) then
+                  begin
+                    href.refaddr:=addr_secrel;
+                    current_asmdata.CurrAsmList.concat(taicpu.op_ref_reg(A_MOV,S_L,href,
+                      newreg(getregtype(reg_offset),getsupreg(reg_offset),R_SUBD)));                       // movl $threadvar@secrel32, %rax
+                  end
+                else // Old binutils (2.21) not support $threadvar@secrel32
+                  begin
+                    href.symbol.typ:=AT_DATA;
+                    href.refaddr:=addr_full;
+                    current_asmdata.CurrAsmList.concat(taicpu.op_ref_reg(A_MOV,S_Q,href,reg_offset));      // movq $threadvar, %rax
+                    reg_start:=cg.getintregister(current_asmdata.CurrAsmList,OS_INT);
+                    reference_reset_symbol(href,current_asmdata.RefAsmSymbol('___tls_start__',AT_DATA),0,sizeof(AInt),[]);
+                    href.refaddr:=addr_full;
+                    current_asmdata.CurrAsmList.concat(taicpu.op_ref_reg(A_MOV,S_Q,href,reg_start));       // movq $___tls_start__, %rdx
+                    current_asmdata.CurrAsmList.concat(taicpu.op_reg_reg(A_SUB,S_Q,reg_start,reg_offset)); // subq %rdx, %rax
+                  end;
+
+                  reference_reset(location.reference,location.reference.alignment,location.reference.volatility);
+                  location.reference.base:=hregister;
+                  location.reference.index:=reg_offset;                                                    // (%rcx,%rax)
+                end;
               else
                 ;
             end;
diff --git a/rtl/inc/heaptrc.pp b/rtl/inc/heaptrc.pp
index 920f448293..da7caade9f 100644
--- a/rtl/inc/heaptrc.pp
+++ b/rtl/inc/heaptrc.pp
@@ -1035,11 +1035,13 @@ var
    edata : ptruint; external name '__data_end__';
    sbss : ptruint; external name '__bss_start__';
    ebss : ptruint; external name '__bss_end__';
+{$ifndef FPC_SECTION_THREADVARS}
    TLSKey : PDWord; external name '_FPC_TlsKey';
    TLSSize : DWord; external name '_FPC_TlsSize';
 
 function TlsGetValue(dwTlsIndex : DWord) : pointer;
   {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'TlsGetValue';
+{$endif !FPC_SECTION_THREADVARS}
 {$endif}
 
 {$ifdef BEOS}
@@ -1104,6 +1106,7 @@ begin
   { inside data, rdata ... bss }
   if (ptruint(p)>=ptruint(@sdata)) and (ptruint(p)<ptruint(@ebss)) then
     exit;
+  {$ifndef FPC_SECTION_THREADVARS}
   { is program multi-threaded and p inside Threadvar range? }
   if TlsKey^<>-1 then
     begin
@@ -1112,6 +1115,7 @@ begin
           (ptruint(p)<ptruint(datap)+TlsSize)) then
         exit;
     end;
+  {$endif !FPC_SECTION_THREADVARS}
 {$endif windows}
 
 {$IFDEF OS2}
diff --git a/rtl/linux/si_impl.inc b/rtl/linux/si_impl.inc
index a727310551..e3ceee427e 100644
--- a/rtl/linux/si_impl.inc
+++ b/rtl/linux/si_impl.inc
@@ -20,7 +20,9 @@ procedure SysEntry_InitTLS(constref info: TEntryInformation); external name 'FPC
 
 var
   InitFinalTable : record end; external name 'INITFINAL';
+  {$ifndef FPC_SECTION_THREADVARS}
   ThreadvarTablesTable : record end; external name 'FPC_THREADVARTABLES';
+  {$endif}
   {$ifdef FPC_HAS_RESSTRINITS}
   ResStrInitTables : record end; external name 'FPC_RESSTRINITTABLES';
   {$endif}
diff --git a/rtl/win/sysosh.inc b/rtl/win/sysosh.inc
index 5bee0d4c17..4b65a1d41d 100644
--- a/rtl/win/sysosh.inc
+++ b/rtl/win/sysosh.inc
@@ -56,7 +56,9 @@ type
     {$ifdef WIN32}
     asm_exit : Procedure;stdcall;
     {$endif WIN32}
+    {$ifndef FPC_SECTION_THREADVARS}
     TlsKeyAddr : PDWord;
+    {$endif !FPC_SECTION_THREADVARS}
     SysInstance: {$ifdef CPU64}PQWord{$else}PLongInt{$endif};
     WideInitTables : Pointer;
   end;
@@ -117,7 +119,4 @@ var
 {$ifdef FPC_USE_WIN64_SEH}
 procedure _fpc_local_unwind(frame,target: Pointer);compilerproc;
 {$endif FPC_USE_WIN64_SEH}
-{$ifdef FPC_SECTION_THREADVARS}
-function fpc_tls_add(addr: pointer):pointer;compilerproc;
-{$endif FPC_SECTION_THREADVARS}
 
diff --git a/rtl/win/systhrd.inc b/rtl/win/systhrd.inc
index 10bb07fda4..dfa583fee7 100644
--- a/rtl/win/systhrd.inc
+++ b/rtl/win/systhrd.inc
@@ -19,6 +19,7 @@
                            Local WINApi imports
 *****************************************************************************}
 
+{$ifndef FPC_SECTION_THREADVARS}
 const
   { LocalAlloc flags  }
   LMEM_FIXED = 0;
@@ -34,15 +35,16 @@ function TlsGetValue(dwTlsIndex : DWord) : pointer;
   {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'TlsGetValue';
 function TlsSetValue(dwTlsIndex : DWord;lpTlsValue : pointer) : LongBool;
   {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'TlsSetValue';
+function LocalAlloc(uFlags:DWord; dwBytes:SIZE_T):Pointer;
+  {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'LocalAlloc';
+function LocalFree(hMem : Pointer):Pointer; {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'LocalFree';
+{$endif !FPC_SECTION_THREADVARS}
 function CreateThread(lpThreadAttributes : pointer;
   dwStackSize : SIZE_T; lpStartAddress : pointer;lpParameter : pointer;
   dwCreationFlags : DWord;var lpThreadId : DWord) : THandle;
   {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'CreateThread';
 procedure ExitThread(dwExitCode : DWord);
   {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'ExitThread';
-function LocalAlloc(uFlags:DWord; dwBytes:SIZE_T):Pointer;
-  {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'LocalAlloc';
-function LocalFree(hMem : Pointer):Pointer; {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'LocalFree';
 procedure Sleep(dwMilliseconds: DWord); {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'Sleep';
 function  WinSuspendThread (threadHandle : THandle) : dword; {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'SuspendThread';
 function  WinResumeThread  (threadHandle : THandle) : dword; {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'ResumeThread';
@@ -98,6 +100,10 @@ var
                              Threadvar support
 *****************************************************************************}
 
+    var
+      MainThreadIdWin32 : DWORD;
+      
+{$ifndef FPC_SECTION_THREADVARS}
     var
       // public names are used by heaptrc unit
       threadvarblocksize : dword; public name '_FPC_TlsSize';
@@ -108,9 +114,6 @@ var
       TLSKey : PDWord = @TLSKeyVar; public name '_FPC_TlsKey';
       {$endif FPC_HAS_INDIRECT_ENTRY_INFORMATION}
 
-    var
-      MainThreadIdWin32 : DWORD;
-
     procedure SysInitThreadvar(var offset : dword;size : dword);
       begin
         offset:=threadvarblocksize;
@@ -199,6 +202,7 @@ var
             TlsSetValue(tlskey^, nil);
           end;
       end;
+{$endif !FPC_SECTION_THREADVARS}
 
 
 {*****************************************************************************
@@ -255,7 +259,9 @@ var
         writeln('Creating new thread');
 {$endif DEBUG_MT}
         { Initialize multithreading if not done }
+{$ifndef FPC_SECTION_THREADVARS}
         SysInitTLS;
+{$endif !FPC_SECTION_THREADVARS}
         IsMultiThread:=true;
 
         { the only way to pass data to the newly created thread
@@ -623,10 +629,17 @@ begin
     EnterCriticalSection   :=@SysEnterCriticalSection;
     TryEnterCriticalSection:=@SysTryEnterCriticalSection;
     LeaveCriticalSection   :=@SysLeaveCriticalSection;
+{$ifndef FPC_SECTION_THREADVARS}
     InitThreadVar          :=@SysInitThreadVar;
     RelocateThreadVar      :=@SysRelocateThreadVar;
     AllocateThreadVars     :=@SysAllocateThreadVars;
     ReleaseThreadVars      :=@SysReleaseThreadVars;
+{$else}
+    InitThreadVar          :=Nil;
+    RelocateThreadVar      :=Nil;
+    AllocateThreadVars     :=Nil;
+    ReleaseThreadVars      :=Nil;
+{$endif FPC_SECTION_THREADVARS}
     BasicEventCreate       :=@intBasicEventCreate;
     BasicEventDestroy      :=@intBasicEventDestroy;
     BasicEventResetEvent   :=@intBasicEventResetEvent;
@@ -641,10 +654,12 @@ begin
     end;
   SetThreadManager(WinThreadManager);
   ThreadID := GetCurrentThreadID;
+{$ifndef FPC_SECTION_THREADVARS}
 {$ifndef FPC_USE_TLS_DIRECTORY}
   if IsLibrary then
 {$endif}
     SysInitTLS;
+{$endif !FPC_SECTION_THREADVARS}
 
 {$ifndef WINCE}
   KernelHandle:=GetModuleHandle(KernelDLL);
diff --git a/rtl/win/systlsdir.inc b/rtl/win/systlsdir.inc
index db53b77a94..d0b48a89a5 100644
--- a/rtl/win/systlsdir.inc
+++ b/rtl/win/systlsdir.inc
@@ -32,8 +32,10 @@ Const
   DLL_PROCESS_DETACH = 0;
   DLL_THREAD_DETACH = 3;
 
+{$ifndef FPC_SECTION_THREADVARS}
 var
    TlsKey : PDWord = @TlsKeyVar;
+{$endif !FPC_SECTION_THREADVARS}
 
 type
   TTlsDirectory=packed record
@@ -44,11 +46,13 @@ type
   end;
 
 
+{$ifndef FPC_SECTION_THREADVARS}
 function TlsGetValue(dwTlsIndex : DWord) : pointer; stdcall;
   external 'kernel32' name 'TlsGetValue';
 
 procedure InitSystemThreads; external name '_FPC_InitSystemThreads';
 procedure SysAllocateThreadVars; external name '_FPC_SysAllocateThreadVars';
+{$endif !FPC_SECTION_THREADVARS}
 procedure InitHeap; external name '_FPC_InitHeap';
 
 {$ifdef FPC_HAS_INDIRECT_ENTRY_INFORMATION}
@@ -58,7 +62,7 @@ procedure SetupEntryInformation(constref info: TEntryInformation); external name
 {$endif FPC_INSSIDE_SYSINIT}
 
 procedure Exec_Tls_callback(Handle : pointer; reason : Dword; Reserved : pointer);
-  stdcall; [public,alias:'_FPC_Tls_Callback'];
+  stdcall;
   begin
      if IsLibrary then
        Exit;
@@ -83,12 +87,19 @@ procedure Exec_Tls_callback(Handle : pointer; reason : Dword; Reserved : pointer
              the threadvar handling can be correctly initialized }
            SetupEntryInformation(SysInitEntryInformation);
            {$endif FPC_HAS_INDIRECT_ENTRY_INFORMATION}
+           { the DLL_PROCESS_ATTACH initialization for the heap is required, because
+             if the DLL_PROCESS_ATTACH entry of a used DLL creates a thread, DLL_THREAD_ATTACH
+             of the executable will be called so that the RTL for the thread can be initialized
+             and that needs the code in DLL_PROCESS_ATTACH to be run first }
            InitHeap;
+         {$ifndef FPC_SECTION_THREADVARS}
            InitSystemThreads;
+         {$endif !FPC_SECTION_THREADVARS}
          end;
 
        DLL_THREAD_ATTACH :
          begin
+         {$ifndef FPC_SECTION_THREADVARS}
          {  !!! SysInitMultithreading must NOT be called here. Windows guarantees that
             the main thread invokes PROCESS_ATTACH, not THREAD_ATTACH. So this always
             executes in non-main thread. SysInitMultithreading() here will cause
@@ -100,6 +111,7 @@ procedure Exec_Tls_callback(Handle : pointer; reason : Dword; Reserved : pointer
          }
            { Allocate Threadvars  }
            SysAllocateThreadVars;
+         {$endif !FPC_SECTION_THREADVARS}
 
            { NS : no idea what is correct to pass here - pass dummy value for now }
            { passing a dummy is ok, the correct value is read from the coff header of SysInstance (FK) }
@@ -107,7 +119,11 @@ procedure Exec_Tls_callback(Handle : pointer; reason : Dword; Reserved : pointer
          end;
        DLL_THREAD_DETACH :
          begin
+         {$ifdef FPC_SECTION_THREADVARS}
+           if ThreadID<>0 then
+         {$else !FPC_SECTION_THREADVARS}
            if TlsGetValue(TLSKey^)<>nil then
+         {$endif !FPC_SECTION_THREADVARS}
              DoneThread; { Assume everything is idempotent there }
          end;
      end;
diff --git a/rtl/win/syswin.inc b/rtl/win/syswin.inc
index 63ee752569..37fb574f33 100644
--- a/rtl/win/syswin.inc
+++ b/rtl/win/syswin.inc
@@ -408,10 +408,12 @@ function Dll_entry{$ifdef FPC_HAS_INDIRECT_ENTRY_INFORMATION}(constref info : TE
          end;
        DLL_THREAD_ATTACH :
          begin
+         {$ifndef FPC_SECTION_THREADVARS}
            { SysInitMultithreading must not be called here,
              see comments in exec_tls_callback below }
            { Allocate Threadvars  }
            SysAllocateThreadVars;
+         {$endif !FPC_SECTION_THREADVARS}
 
            { NS : no idea what is correct to pass here - pass dummy value for now }
            { passing a dummy is ok, the correct value is read from the coff header of SysInstance (FK) }
@@ -425,7 +427,11 @@ function Dll_entry{$ifdef FPC_HAS_INDIRECT_ENTRY_INFORMATION}(constref info : TE
            if assigned(Dll_Thread_Detach_Hook) then
              Dll_Thread_Detach_Hook(DllParam);
            { Release Threadvars }
+         {$ifdef FPC_SECTION_THREADVARS}
+           if ThreadID<>0 then
+         {$else !FPC_SECTION_THREADVARS}
            if TlsGetValue(TLSKey^)<>nil then
+         {$endif !FPC_SECTION_THREADVARS}
              DoneThread; { Assume everything is idempotent there }
          end;
        DLL_PROCESS_DETACH :
@@ -439,9 +445,11 @@ function Dll_entry{$ifdef FPC_HAS_INDIRECT_ENTRY_INFORMATION}(constref info : TE
                InternalExit;
              end;
 
+         {$ifndef FPC_SECTION_THREADVARS}
            SysReleaseThreadVars;
            { Free TLS resources used by ThreadVars }
            SysFiniMultiThreading;
+         {$endif !FPC_SECTION_THREADVARS}
            MainThreadIDWin32:=0;
          end;
      end;
diff --git a/rtl/win32/sysinit.inc b/rtl/win32/sysinit.inc
index 0eea501af0..c8f7b6f042 100644
--- a/rtl/win32/sysinit.inc
+++ b/rtl/win32/sysinit.inc
@@ -16,10 +16,14 @@
 
    var
       SysInstance : LongInt;
+ {$ifndef FPC_SECTION_THREADVARS}
       TlsKeyVar: DWord = $ffffffff;
+ {$endif !FPC_SECTION_THREADVARS}
 
       InitFinalTable : record end; external name 'INITFINAL';
+ {$ifndef FPC_SECTION_THREADVARS}
       ThreadvarTablesTable : record end; external name 'FPC_THREADVARTABLES';
+ {$endif !FPC_SECTION_THREADVARS}
       WideInitTables : record end; external name 'FPC_WIDEINITTABLES';
  {$ifdef FPC_HAS_RESSTRINITS}
       ResStrInitTables : record end; external name 'FPC_RESSTRINITTABLES';
@@ -53,7 +57,11 @@
       STD_INPUT_HANDLE = dword(-10);
       SysInitEntryInformation : TEntryInformation = (
         InitFinalTable : @InitFinalTable;
+{$ifdef FPC_SECTION_THREADVARS}
+        ThreadvarTablesTable : nil;
+{$else FPC_SECTION_THREADVARS}
         ThreadvarTablesTable : @ThreadvarTablesTable;
+{$endif FPC_SECTION_THREADVARS}
         ResourceStringTables : @ResourceStringTables;
 {$ifdef FPC_HAS_RESSTRINITS}
         ResStrInitTables : @ResStrInitTables;
@@ -65,7 +73,9 @@
         valgrind_used : false;
         OS : (
           asm_exit : @asm_exit;
+{$ifndef FPC_SECTION_THREADVARS}
           TlsKeyAddr : @TlsKeyVar;
+{$endif !FPC_SECTION_THREADVARS}
           SysInstance : @SysInstance;
           WideInitTables: @WideInitTables;
           );
diff --git a/rtl/win32/system.pp b/rtl/win32/system.pp
index 4ffa0b65ff..4cf69de94d 100644
--- a/rtl/win32/system.pp
+++ b/rtl/win32/system.pp
@@ -78,7 +78,9 @@ end;
 
 procedure OsSetupEntryInformation(constref info: TEntryInformation);
 begin
+{$ifndef FPC_SECTION_THREADVARS}
   TlsKey := info.OS.TlsKeyAddr;
+{$endif !FPC_SECTION_THREADVARS}
   FPCSysInstance := info.OS.SysInstance;
   WStrInitTablesTable := info.OS.WideInitTables;
 end;
@@ -498,41 +500,6 @@ begin
 end;
 {$endif Set_i386_Exception_handler}
 
-{$ifdef FPC_SECTION_THREADVARS}
-function fpc_tls_add(addr: pointer): pointer; assembler; nostackframe;
-  [public,alias: 'FPC_TLS_ADD']; compilerproc;
-  asm
-      sub   $tls_data_start,%eax
-      cmpb  $0,IsLibrary
-      mov   _tls_index,%ecx
-      jnz   .L1
-      mov   %fs:(0x2c),%edx
-      add   (%edx,%ecx,4),%eax
-      ret
-.L1:
-      push  %ebx
-      mov   %eax,%ebx
-      call  GetLastError
-      push  %eax                      { save LastError }
-      push  _tls_index
-      call  TlsGetValue
-      test  %eax,%eax
-      jnz   .L2
-      { This can happen when a thread existed before DLL was loaded,
-        or if DisableThreadLibraryCalls was called. }
-      call  SysAllocateThreadVars
-      mov   $0x1000000,%eax
-      call  InitThread
-      push  _tls_index
-      call  TlsGetValue
-.L2:
-      add   %eax,%ebx
-      call  SetLastError              { restore (value is on stack) }
-      mov   %ebx,%eax
-      pop   %ebx
-  end;
-{$endif FPC_SECTION_THREADVARS}
-
 function CheckInitialStkLen(stklen : SizeUInt) : SizeUInt;
   type
     tdosheader = packed record
diff --git a/rtl/win64/sysinit.pp b/rtl/win64/sysinit.pp
index 5d396beb8f..9d740e19ec 100644
--- a/rtl/win64/sysinit.pp
+++ b/rtl/win64/sysinit.pp
@@ -21,10 +21,14 @@ unit sysinit;
 
    var
       SysInstance : QWord;
+{$ifndef FPC_SECTION_THREADVARS}
       TlsKeyVar: DWord = $ffffffff;
+{$endif !FPC_SECTION_THREADVARS}
 
       InitFinalTable : record end; external name 'INITFINAL';
+{$ifndef FPC_SECTION_THREADVARS}
       ThreadvarTablesTable : record end; external name 'FPC_THREADVARTABLES';
+{$endif !FPC_SECTION_THREADVARS}
       WideInitTables : record end; external name 'FPC_WIDEINITTABLES';
 {$ifdef FPC_HAS_RESSTRINITS}
       ResStrInitTables : record end; external name 'FPC_RESSTRINITTABLES';
@@ -58,7 +62,11 @@ unit sysinit;
       STD_INPUT_HANDLE = dword(-10);
       SysInitEntryInformation : TEntryInformation = (
         InitFinalTable : @InitFinalTable;
+{$ifdef FPC_SECTION_THREADVARS}
+        ThreadvarTablesTable : nil;
+{$else FPC_SECTION_THREADVARS}
         ThreadvarTablesTable : @ThreadvarTablesTable;
+{$endif FPC_SECTION_THREADVARS}
         ResourceStringTables : @ResourceStringTables;
 {$ifdef FPC_HAS_RESSTRINITS}
         ResStrInitTables : @ResStrInitTables;
@@ -69,7 +77,9 @@ unit sysinit;
         PascalMain : @PascalMain;
         valgrind_used : false;
         OS : (
+{$ifndef FPC_SECTION_THREADVARS}
           TlsKeyAddr : @TlsKeyVar;
+{$endif !FPC_SECTION_THREADVARS}
           SysInstance : @SysInstance;
           WideInitTables: @WideInitTables;
           );
@@ -124,6 +134,9 @@ unit sysinit;
       sysinstance:=_hinstance;
       dllreason:=_dllreason;
       dllparam:=PtrInt(_dllparam);
+{$if defined(FPC_USE_TLS_DIRECTORY) and defined(FPC_SECTION_THREADVARS)}
+      LinkIn(@tlsdir,@tls_callback_end,@tls_callback);
+{$endif}
       SetupEntryInformation;
       DLL_Entry(SysInitEntryInformation);
     end;
@@ -135,6 +148,9 @@ unit sysinit;
       sysinstance:=_hinstance;
       dllreason:=_dllreason;
       dllparam:=PtrInt(_dllparam);
+{$if defined(FPC_USE_TLS_DIRECTORY) and defined(FPC_SECTION_THREADVARS)}
+      LinkIn(@tlsdir,@tls_callback_end,@tls_callback);
+{$endif}
       SetupEntryInformation;
       DLL_Entry(SysInitEntryInformation);
     end;
diff --git a/rtl/win64/system.pp b/rtl/win64/system.pp
index 6cf7b174cd..b6fad06fff 100644
--- a/rtl/win64/system.pp
+++ b/rtl/win64/system.pp
@@ -122,7 +122,9 @@ procedure PascalMain;external name 'PASCALMAIN';
 
 procedure OsSetupEntryInformation(constref info: TEntryInformation);
 begin
+{$ifndef FPC_SECTION_THREADVARS}
   TlsKey := info.OS.TlsKeyAddr;
+{$endif !FPC_SECTION_THREADVARS}
   FPCSysInstance := info.OS.SysInstance;
   WStrInitTablesTable := info.OS.WideInitTables;
 end;
@@ -504,47 +506,6 @@ begin
 end;
 {$endif VER3_0}
 
-{$ifdef FPC_SECTION_THREADVARS}
-function fpc_tls_add(addr: pointer): pointer; assembler; nostackframe;
-  [public,alias: 'FPC_TLS_ADD']; compilerproc;
-  asm
-      sub   $56,%rsp                  { 32 spill area + 16 local vars + 8 misalignment }
-  .seh_stackalloc 56
-  .seh_endprologue
-      lea   tls_data_start(%rip),%rax
-      sub   %rax,%rcx
-      cmpb  $0,IsLibrary(%rip)
-      mov   _tls_index(%rip),%eax
-      jnz   .L1
-      mov   %gs:(88),%rdx
-      add   (%rdx,%rax,8),%rcx
-      mov   %rcx,%rax
-      jmp   .L3
-.L1:
-      mov   %rcx,32(%rsp)
-      call  GetLastError
-      mov   %rax,40(%rsp)             { save LastError }
-      mov   _tls_index(%rip),%ecx
-      call  TlsGetValue
-      test  %rax,%rax
-      jnz   .L2
-      { This can happen when a thread existed before DLL was loaded,
-        or if DisableThreadLibraryCalls was called. }
-      call  SysAllocateThreadVars
-      mov   $0x1000000,%rcx
-      call  InitThread
-      mov   _tls_index(%rip),%ecx
-      call  TlsGetValue
-.L2:
-      add   %rax,32(%rsp)
-      mov   40(%rsp),%rcx
-      call  SetLastError
-      mov   32(%rsp),%rax
-.L3:
-      add   $56,%rsp
-  end;
-{$endif FPC_SECTION_THREADVARS}
-
 function CheckInitialStkLen(stklen : SizeUInt) : SizeUInt;
   type
     tdosheader = packed record
-- 
2.26.2.windows.1

---
 compiler/aggas.pas         |  4 +-
 compiler/cgbase.pas        |  1 +
 compiler/dbgdwarf.pas      |  3 +-
 compiler/nld.pas           |  4 +-
 compiler/nutils.pas        |  9 +++--
 compiler/ogcoff.pas        | 35 +++++++----------
 compiler/systems/i_win.pas |  3 ++
 compiler/systems/t_win.pas | 27 ++++++++++++-
 compiler/x86/aasmcpu.pas   |  6 +++
 compiler/x86/agx86att.pas  |  9 +++++
 compiler/x86/nx86ld.pas    | 79 +++++++++++++++++++++++---------------
 rtl/inc/heaptrc.pp         |  4 ++
 rtl/linux/si_impl.inc      |  2 +
 rtl/win/sysosh.inc         |  5 +--
 rtl/win/systhrd.inc        | 27 ++++++++++---
 rtl/win/systlsdir.inc      | 21 ++++++++--
 rtl/win/syswin.inc         |  8 ++++
 rtl/win32/sysinit.inc      | 10 +++++
 rtl/win32/system.pp        | 37 +-----------------
 rtl/win64/sysinit.pp       | 16 ++++++++
 rtl/win64/system.pp        | 43 +--------------------
 21 files changed, 204 insertions(+), 149 deletions(-)

diff --git a/compiler/aggas.pas b/compiler/aggas.pas
index 41d20c5129..f9b69f605b 100644
--- a/compiler/aggas.pas
+++ b/compiler/aggas.pas
@@ -533,7 +533,9 @@ implementation
              { according to the GNU AS guide AS for COFF does not support the
                progbits }
              writer.AsmWrite('.section ');
-             usesectionflags:=true;
+             { .tls section requires special processing }
+             if not((atype=sec_threadvar) and (tf_section_threadvars in target_info.flags)) then
+               usesectionflags:=true;
            end;
          system_powerpc_darwin,
          system_i386_darwin,
diff --git a/compiler/cgbase.pas b/compiler/cgbase.pas
index 6dca71ba40..0e862ca739 100644
--- a/compiler/cgbase.pas
+++ b/compiler/cgbase.pas
@@ -140,6 +140,7 @@ interface
 {$ifdef x86_64}
           ,addr_tpoff
           ,addr_tlsgd
+          ,addr_secrel
 {$endif x86_64}
          );
 
diff --git a/compiler/dbgdwarf.pas b/compiler/dbgdwarf.pas
index 5b75e257b5..d81436ed2a 100644
--- a/compiler/dbgdwarf.pas
+++ b/compiler/dbgdwarf.pas
@@ -2658,7 +2658,8 @@ implementation
                   begin
                     if vo_is_thread_var in sym.varoptions then
                       begin
-                        if tf_section_threadvars in target_info.flags then
+                        if (tf_section_threadvars in target_info.flags) and
+                           (not (target_info.system in [system_x86_64_win64])) then // TODO : dwarf for threadvars on win64
                           begin
                             case sizeof(puint) of
                               2:
diff --git a/compiler/nld.pas b/compiler/nld.pas
index e2bccbc080..0453f162e8 100644
--- a/compiler/nld.pas
+++ b/compiler/nld.pas
@@ -450,7 +450,9 @@ implementation
                 { call to get address of threadvar }
                 if (vo_is_thread_var in tabstractvarsym(symtableentry).varoptions) then
                   begin
-                    include(current_procinfo.flags,pi_do_call);
+                    if (not (tf_section_threadvars in target_info.flags)) or
+                       (current_settings.tlsmodel in [tlsm_local_dynamic,tlsm_global_dynamic]) then
+                      include(current_procinfo.flags,pi_do_call);
                     include(current_procinfo.flags,pi_uses_threadvar);
                   end;
               end;
diff --git a/compiler/nutils.pas b/compiler/nutils.pas
index d586fc222d..b5c0de1937 100644
--- a/compiler/nutils.pas
+++ b/compiler/nutils.pas
@@ -191,7 +191,7 @@ interface
 implementation
 
     uses
-      cutils,verbose,globals,
+      cutils,verbose,globals,systems,
       symconst,symdef,
       defcmp,defutil,
       ncon,ncnv,nld,nflw,nset,ncal,nadd,nmem,ninl,
@@ -799,10 +799,13 @@ implementation
                 begin
                   if assigned(tloadnode(p).left) then
                     inc(result,node_complexity(tloadnode(p).left));
-                  { threadvars need a helper call }
+                  { some threadvars models need a helper call }
                   if (tloadnode(p).symtableentry.typ=staticvarsym) and
                      (vo_is_thread_var in tstaticvarsym(tloadnode(p).symtableentry).varoptions) then
-                    inc(result,5)
+                    if (tf_section_threadvars in target_info.flags) and (current_settings.tlsmodel in [tlsm_local_exec,tlsm_initial_exec]) then
+                      inc(result,3)
+                    else
+                      inc(result,50)
                   else if not((tloadnode(p).symtableentry.typ in [staticvarsym,localvarsym,paravarsym,fieldvarsym]) and
                     (tabstractvarsym(tloadnode(p).symtableentry).varregable in [vr_intreg,vr_mmreg,vr_fpureg])) then
                     inc(result);
diff --git a/compiler/ogcoff.pas b/compiler/ogcoff.pas
index 8c05b3792a..90bbca8b0d 100644
--- a/compiler/ogcoff.pas
+++ b/compiler/ogcoff.pas
@@ -1015,8 +1015,10 @@ const pemagic : array[0..3] of byte = (
                   end;
                 RELOC_SECREL32 :
                   begin
-                    { fixup address when the symbol was known in defined object }
-                    if (relocsec.objdata=objsec.objdata) then
+                    { fixup address when the symbol was known in defined object or
+                      when the symbol is a threadvar }
+                    if (relocsec.objdata=objsec.objdata) or
+                       assigned(objreloc.symbol) and (copy(objreloc.symbol.objsection.Name,1,4)='.tls') then
                       dec(address,relocsec.ExeSection.MemPos);
                     inc(address,relocval);
                   end;
@@ -2667,21 +2669,17 @@ const pemagic : array[0..3] of byte = (
 
         procedure UpdateTlsDataDir;
         var
-          {callbacksection : TExeSection;}
           tlsexesymbol: TExeSymbol;
           tlssymbol: TObjSymbol;
-          callbackexesymbol: TExeSymbol;
-          //callbacksymbol: TObjSymbol;
         begin
           { according to GNU ld,
             the callback routines should be placed into .CRT$XL*
             sections, and the thread local variables in .tls
             __tls_start__ and __tls_end__ symbols
-            should be used for the initialized part,
-            which we do not support yet. }
-          { For now, we only pass the address of the __tls_used
+            should be used for the initialized part. }
+          { We pass the address of the __tls_used
             asm symbol into PE_DATADIR_TLS with the correct
-            size of this table (different for win32/win64 }
+            size of this table (different for win32/win64) }
           tlsexesymbol:=texesymbol(ExeSymbolList.Find(
             target_info.Cprefix+'_tls_used'));
           if assigned(tlsexesymbol) then
@@ -2690,18 +2688,6 @@ const pemagic : array[0..3] of byte = (
               peoptheader.DataDirectory[PE_DATADIR_TLS].vaddr:=tlssymbol.address;
               { sizeof(TlsDirectory) is different on host and target when cross-compiling }
               peoptheader.DataDirectory[PE_DATADIR_TLS].size:=TLSDIR_SIZE;
-              if IsSharedLibrary then
-                begin
-                  { Here we should reset __FPC_tls_callbacks value to nil }
-                  callbackexesymbol:=texesymbol(ExeSymbolList.Find(
-                                        '__FPC_tls_callbacks'));
-                  if assigned (callbackexesymbol) then
-                    begin
-                      //callbacksymbol:=callbackexesymbol.ObjSymbol;
-
-                    end;
-                end;
-
            end;
         end;
 
@@ -2790,7 +2776,10 @@ const pemagic : array[0..3] of byte = (
               end
             else
               begin
-                peoptheader.MajorOperatingSystemVersion:=4;
+                if tf_section_threadvars in target_info.flags then
+                  peoptheader.MajorOperatingSystemVersion:=6 // >= Windows Vista
+                else
+                  peoptheader.MajorOperatingSystemVersion:=4;
                 peoptheader.MinorOperatingSystemVersion:=0;
               end;
             if SetPEUserVersionSetExplicitely then
@@ -2812,6 +2801,8 @@ const pemagic : array[0..3] of byte = (
               begin
                 if target_info.system in systems_wince then
                   peoptheader.MajorSubsystemVersion:=3
+                else if tf_section_threadvars in target_info.flags then
+                  peoptheader.MajorSubsystemVersion:=6 // >= Windows Vista
                 else
                   peoptheader.MajorSubsystemVersion:=4;
                 peoptheader.MinorSubsystemVersion:=0;
diff --git a/compiler/systems/i_win.pas b/compiler/systems/i_win.pas
index d4e8f842d5..d1ffb4e57c 100644
--- a/compiler/systems/i_win.pas
+++ b/compiler/systems/i_win.pas
@@ -107,6 +107,9 @@ unit i_win;
             shortname    : 'Win64';
             flags        : [tf_files_case_aware,tf_has_dllscanner,
                             tf_smartlink_sections,
+{$if not defined(DISABLE_TLS_DIRECTORY) and not defined(SUPPORT_WINXP)}
+                            tf_section_threadvars,
+{$endif}
                             tf_winlikewidestring,tf_no_pic_supported,
                             tf_dwarf_only_local_labels,
                             tf_no_generic_stackcheck,tf_has_winlike_resources,
diff --git a/compiler/systems/t_win.pas b/compiler/systems/t_win.pas
index 887787f8e9..b96f820a3c 100644
--- a/compiler/systems/t_win.pas
+++ b/compiler/systems/t_win.pas
@@ -1711,11 +1711,36 @@ implementation
             else
               ;
           end;
-        if dllversion<>'' then
+        if SetPEUserVersionSetExplicitely then
+          begin
+            peoptheader.MajorImageVersion:=peuserversionmajor;
+            peoptheader.MinorImageVersion:=peuserversionminor;
+          end
+        else if dllversion<>'' then
           begin
            peoptheader.MajorImageVersion:=dllmajor;
            peoptheader.MinorImageVersion:=dllminor;
           end;
+        if SetPEOSVersionSetExplicitely then
+          begin
+            peoptheader.MajorOperatingSystemVersion:=peosversionmajor;
+            peoptheader.MinorOperatingSystemVersion:=peosversionminor;
+          end
+        else if tf_section_threadvars in target_info.flags then
+          begin
+            peoptheader.MajorOperatingSystemVersion:=6; // >= Windows Vista
+            peoptheader.MinorOperatingSystemVersion:=0;
+          end;
+        if SetPESubSysVersionSetExplicitely then
+          begin
+            peoptheader.MajorSubsystemVersion:=pesubsysversionmajor;
+            peoptheader.MinorSubsystemVersion:=pesubsysversionminor;
+          end
+        else if tf_section_threadvars in target_info.flags then
+          begin
+            peoptheader.MajorSubsystemVersion:=6; // >= Windows Vista
+            peoptheader.MinorSubsystemVersion:=0;
+          end;
         { reset timestamp }
         peheader.time:=0;
         { write header back, skip pe magic }
diff --git a/compiler/x86/aasmcpu.pas b/compiler/x86/aasmcpu.pas
index f712af7196..115474e6d1 100644
--- a/compiler/x86/aasmcpu.pas
+++ b/compiler/x86/aasmcpu.pas
@@ -3617,6 +3617,12 @@ implementation
                       currabsreloc:=RELOC_TLSGD;
                       currabsreloc32:=RELOC_TLSGD;
                     end
+                  else if oper[opidx]^.ref^.refaddr=addr_secrel then
+                    begin
+                      currrelreloc:=RELOC_SECREL32;
+                      currabsreloc:=RELOC_SECREL32;
+                      currabsreloc32:=RELOC_SECREL32;
+                    end
                   else
 {$endif x86_64}
                     begin
diff --git a/compiler/x86/agx86att.pas b/compiler/x86/agx86att.pas
index 657f073163..bf69e08bd5 100644
--- a/compiler/x86/agx86att.pas
+++ b/compiler/x86/agx86att.pas
@@ -190,6 +190,8 @@ interface
                owner.writer.AsmWrite('@tpoff');
              addr_tlsgd:
                owner.writer.AsmWrite('@tlsgd');
+             addr_secrel:
+               owner.writer.AsmWrite('@secrel32');
 {$endif x86_64}
              else
                ;
@@ -259,6 +261,13 @@ interface
               {$ifdef x86_64},addr_tpoff,addr_tlsgd{$endif x86_64}
               ] then
               WriteReference(o.ref^)
+        {$ifdef x86_64}
+            else if o.ref^.refaddr=addr_secrel then
+              begin
+                owner.writer.AsmWrite('$');
+                WriteReference(o.ref^);
+              end
+        {$endif x86_64}
             else
               begin
                 owner.writer.AsmWrite('$');
diff --git a/compiler/x86/nx86ld.pas b/compiler/x86/nx86ld.pas
index 774addd1ca..61c82755d4 100644
--- a/compiler/x86/nx86ld.pas
+++ b/compiler/x86/nx86ld.pas
@@ -38,10 +38,10 @@ interface
 implementation
 
     uses
-      globals,
+      globals,fmodule,
       cutils,verbose,systems,
       aasmbase,aasmtai,aasmdata,aasmcpu,
-      cgutils,cgobj,
+      cgutils,cgobj,hlcgobj,
       symconst,symdef,symtable,
       cgbase,cpubase,parabase,paramgr,
       procinfo;
@@ -54,37 +54,10 @@ implementation
       var
         paraloc1 : tcgpara;
         pd: tprocdef;
-        href: treference;
-        hregister : tregister;
-        handled: boolean;
+        href, href_index: treference;
+        hregister, reg_index, reg_offset, reg_start : tregister;
       begin
-        handled:=false;
-        if (tf_section_threadvars in target_info.flags) then
-          begin
-            if target_info.system in [system_i386_win32,system_x86_64_win64] then
-              begin
-                paraloc1.init;
-                pd:=search_system_proc('fpc_tls_add');
-                paramanager.getcgtempparaloc(current_asmdata.CurrAsmList,pd,1,paraloc1);
-                if not(vo_is_weak_external in gvs.varoptions) then
-                  reference_reset_symbol(href,current_asmdata.RefAsmSymbol(gvs.mangledname,AT_DATA,use_indirect_symbol(gvs)),0,sizeof(pint),[])
-                else
-                  reference_reset_symbol(href,current_asmdata.WeakRefAsmSymbol(gvs.mangledname,AT_DATA),0,sizeof(pint),[]);
-                cg.a_loadaddr_ref_cgpara(current_asmdata.CurrAsmList,href,paraloc1);
-                paramanager.freecgpara(current_asmdata.CurrAsmList,paraloc1);
-                paraloc1.done;
-
-                cg.g_call(current_asmdata.CurrAsmList,'FPC_TLS_ADD');
-                cg.ungetcpuregister(current_asmdata.CurrAsmList,NR_FUNCTION_RESULT_REG);
-                hregister:=cg.getaddressregister(current_asmdata.CurrAsmList);
-                cg.a_load_reg_reg(current_asmdata.CurrAsmList,OS_ADDR,OS_ADDR,NR_FUNCTION_RESULT_REG,hregister);
-                location.reference.base:=hregister;
-                handled:=true;
-              end;
-          end;
-
-        if not handled then
-          inherited;
+        inherited;
 
         if (tf_section_threadvars in target_info.flags) then
           begin
@@ -156,6 +129,48 @@ implementation
                       Internalerror(2019012002);
                   end;
                 end;
+              system_x86_64_win64:
+                begin
+                  reference_reset(href,sizeof(AInt),[]); href.segment:=NR_GS; href.offset:=88;
+                  hregister:=cg.getaddressregister(current_asmdata.CurrAsmList);
+                  current_asmdata.CurrAsmList.concat(taicpu.op_ref_reg(A_MOV,S_Q,href,hregister));         // movq %gs:88, %rcx
+
+                  reference_reset_base(href,hregister,0,ctempposinvalid,sizeof(AInt),[]);
+                  if (main_module=nil) or main_module.IsLibrary or main_module.is_unit then
+                    begin // _tls_index <> 0
+                      reference_reset_symbol(href_index,current_asmdata.RefAsmSymbol('_tls_index',AT_DATA),0,sizeof(dword),[]);
+                      reg_index:=cg.getintregister(current_asmdata.CurrAsmList,OS_INT);
+                      hlcg.a_load_ref_reg(current_asmdata.CurrAsmList,u32inttype,u64inttype,
+                        href_index,reg_index);                                                             // movl _tls_index(%rip), %edx
+                      href.index:=reg_index; href.scalefactor:=8;
+                    end;
+                  current_asmdata.CurrAsmList.concat(taicpu.op_ref_reg(A_MOV,S_Q,href,hregister));         // movq (%rcx), %rcx  |  movq (%rcx,%rdx,8), %rcx
+
+                  reg_offset:=cg.getintregister(current_asmdata.CurrAsmList,OS_INT);
+                  href:=location.reference; // from inherited
+
+                if not (cs_asm_pre_binutils_2_25 in current_settings.globalswitches) then
+                  begin
+                    href.refaddr:=addr_secrel;
+                    current_asmdata.CurrAsmList.concat(taicpu.op_ref_reg(A_MOV,S_L,href,
+                      newreg(getregtype(reg_offset),getsupreg(reg_offset),R_SUBD)));                       // movl $threadvar@secrel32, %rax
+                  end
+                else // Old binutils (2.21) not support $threadvar@secrel32
+                  begin
+                    href.symbol.typ:=AT_DATA;
+                    href.refaddr:=addr_full;
+                    current_asmdata.CurrAsmList.concat(taicpu.op_ref_reg(A_MOV,S_Q,href,reg_offset));      // movq $threadvar, %rax
+                    reg_start:=cg.getintregister(current_asmdata.CurrAsmList,OS_INT);
+                    reference_reset_symbol(href,current_asmdata.RefAsmSymbol('___tls_start__',AT_DATA),0,sizeof(AInt),[]);
+                    href.refaddr:=addr_full;
+                    current_asmdata.CurrAsmList.concat(taicpu.op_ref_reg(A_MOV,S_Q,href,reg_start));       // movq $___tls_start__, %rdx
+                    current_asmdata.CurrAsmList.concat(taicpu.op_reg_reg(A_SUB,S_Q,reg_start,reg_offset)); // subq %rdx, %rax
+                  end;
+
+                  reference_reset(location.reference,location.reference.alignment,location.reference.volatility);
+                  location.reference.base:=hregister;
+                  location.reference.index:=reg_offset;                                                    // (%rcx,%rax)
+                end;
               else
                 ;
             end;
diff --git a/rtl/inc/heaptrc.pp b/rtl/inc/heaptrc.pp
index 920f448293..da7caade9f 100644
--- a/rtl/inc/heaptrc.pp
+++ b/rtl/inc/heaptrc.pp
@@ -1035,11 +1035,13 @@ var
    edata : ptruint; external name '__data_end__';
    sbss : ptruint; external name '__bss_start__';
    ebss : ptruint; external name '__bss_end__';
+{$ifndef FPC_SECTION_THREADVARS}
    TLSKey : PDWord; external name '_FPC_TlsKey';
    TLSSize : DWord; external name '_FPC_TlsSize';
 
 function TlsGetValue(dwTlsIndex : DWord) : pointer;
   {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'TlsGetValue';
+{$endif !FPC_SECTION_THREADVARS}
 {$endif}
 
 {$ifdef BEOS}
@@ -1104,6 +1106,7 @@ begin
   { inside data, rdata ... bss }
   if (ptruint(p)>=ptruint(@sdata)) and (ptruint(p)<ptruint(@ebss)) then
     exit;
+  {$ifndef FPC_SECTION_THREADVARS}
   { is program multi-threaded and p inside Threadvar range? }
   if TlsKey^<>-1 then
     begin
@@ -1112,6 +1115,7 @@ begin
           (ptruint(p)<ptruint(datap)+TlsSize)) then
         exit;
     end;
+  {$endif !FPC_SECTION_THREADVARS}
 {$endif windows}
 
 {$IFDEF OS2}
diff --git a/rtl/linux/si_impl.inc b/rtl/linux/si_impl.inc
index a727310551..e3ceee427e 100644
--- a/rtl/linux/si_impl.inc
+++ b/rtl/linux/si_impl.inc
@@ -20,7 +20,9 @@ procedure SysEntry_InitTLS(constref info: TEntryInformation); external name 'FPC
 
 var
   InitFinalTable : record end; external name 'INITFINAL';
+  {$ifndef FPC_SECTION_THREADVARS}
   ThreadvarTablesTable : record end; external name 'FPC_THREADVARTABLES';
+  {$endif}
   {$ifdef FPC_HAS_RESSTRINITS}
   ResStrInitTables : record end; external name 'FPC_RESSTRINITTABLES';
   {$endif}
diff --git a/rtl/win/sysosh.inc b/rtl/win/sysosh.inc
index 5bee0d4c17..4b65a1d41d 100644
--- a/rtl/win/sysosh.inc
+++ b/rtl/win/sysosh.inc
@@ -56,7 +56,9 @@ type
     {$ifdef WIN32}
     asm_exit : Procedure;stdcall;
     {$endif WIN32}
+    {$ifndef FPC_SECTION_THREADVARS}
     TlsKeyAddr : PDWord;
+    {$endif !FPC_SECTION_THREADVARS}
     SysInstance: {$ifdef CPU64}PQWord{$else}PLongInt{$endif};
     WideInitTables : Pointer;
   end;
@@ -117,7 +119,4 @@ var
 {$ifdef FPC_USE_WIN64_SEH}
 procedure _fpc_local_unwind(frame,target: Pointer);compilerproc;
 {$endif FPC_USE_WIN64_SEH}
-{$ifdef FPC_SECTION_THREADVARS}
-function fpc_tls_add(addr: pointer):pointer;compilerproc;
-{$endif FPC_SECTION_THREADVARS}
 
diff --git a/rtl/win/systhrd.inc b/rtl/win/systhrd.inc
index 10bb07fda4..dfa583fee7 100644
--- a/rtl/win/systhrd.inc
+++ b/rtl/win/systhrd.inc
@@ -19,6 +19,7 @@
                            Local WINApi imports
 *****************************************************************************}
 
+{$ifndef FPC_SECTION_THREADVARS}
 const
   { LocalAlloc flags  }
   LMEM_FIXED = 0;
@@ -34,15 +35,16 @@ function TlsGetValue(dwTlsIndex : DWord) : pointer;
   {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'TlsGetValue';
 function TlsSetValue(dwTlsIndex : DWord;lpTlsValue : pointer) : LongBool;
   {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'TlsSetValue';
+function LocalAlloc(uFlags:DWord; dwBytes:SIZE_T):Pointer;
+  {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'LocalAlloc';
+function LocalFree(hMem : Pointer):Pointer; {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'LocalFree';
+{$endif !FPC_SECTION_THREADVARS}
 function CreateThread(lpThreadAttributes : pointer;
   dwStackSize : SIZE_T; lpStartAddress : pointer;lpParameter : pointer;
   dwCreationFlags : DWord;var lpThreadId : DWord) : THandle;
   {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'CreateThread';
 procedure ExitThread(dwExitCode : DWord);
   {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'ExitThread';
-function LocalAlloc(uFlags:DWord; dwBytes:SIZE_T):Pointer;
-  {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'LocalAlloc';
-function LocalFree(hMem : Pointer):Pointer; {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'LocalFree';
 procedure Sleep(dwMilliseconds: DWord); {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'Sleep';
 function  WinSuspendThread (threadHandle : THandle) : dword; {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'SuspendThread';
 function  WinResumeThread  (threadHandle : THandle) : dword; {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'ResumeThread';
@@ -98,6 +100,10 @@ var
                              Threadvar support
 *****************************************************************************}
 
+    var
+      MainThreadIdWin32 : DWORD;
+      
+{$ifndef FPC_SECTION_THREADVARS}
     var
       // public names are used by heaptrc unit
       threadvarblocksize : dword; public name '_FPC_TlsSize';
@@ -108,9 +114,6 @@ var
       TLSKey : PDWord = @TLSKeyVar; public name '_FPC_TlsKey';
       {$endif FPC_HAS_INDIRECT_ENTRY_INFORMATION}
 
-    var
-      MainThreadIdWin32 : DWORD;
-
     procedure SysInitThreadvar(var offset : dword;size : dword);
       begin
         offset:=threadvarblocksize;
@@ -199,6 +202,7 @@ var
             TlsSetValue(tlskey^, nil);
           end;
       end;
+{$endif !FPC_SECTION_THREADVARS}
 
 
 {*****************************************************************************
@@ -255,7 +259,9 @@ var
         writeln('Creating new thread');
 {$endif DEBUG_MT}
         { Initialize multithreading if not done }
+{$ifndef FPC_SECTION_THREADVARS}
         SysInitTLS;
+{$endif !FPC_SECTION_THREADVARS}
         IsMultiThread:=true;
 
         { the only way to pass data to the newly created thread
@@ -623,10 +629,17 @@ begin
     EnterCriticalSection   :=@SysEnterCriticalSection;
     TryEnterCriticalSection:=@SysTryEnterCriticalSection;
     LeaveCriticalSection   :=@SysLeaveCriticalSection;
+{$ifndef FPC_SECTION_THREADVARS}
     InitThreadVar          :=@SysInitThreadVar;
     RelocateThreadVar      :=@SysRelocateThreadVar;
     AllocateThreadVars     :=@SysAllocateThreadVars;
     ReleaseThreadVars      :=@SysReleaseThreadVars;
+{$else}
+    InitThreadVar          :=Nil;
+    RelocateThreadVar      :=Nil;
+    AllocateThreadVars     :=Nil;
+    ReleaseThreadVars      :=Nil;
+{$endif FPC_SECTION_THREADVARS}
     BasicEventCreate       :=@intBasicEventCreate;
     BasicEventDestroy      :=@intBasicEventDestroy;
     BasicEventResetEvent   :=@intBasicEventResetEvent;
@@ -641,10 +654,12 @@ begin
     end;
   SetThreadManager(WinThreadManager);
   ThreadID := GetCurrentThreadID;
+{$ifndef FPC_SECTION_THREADVARS}
 {$ifndef FPC_USE_TLS_DIRECTORY}
   if IsLibrary then
 {$endif}
     SysInitTLS;
+{$endif !FPC_SECTION_THREADVARS}
 
 {$ifndef WINCE}
   KernelHandle:=GetModuleHandle(KernelDLL);
diff --git a/rtl/win/systlsdir.inc b/rtl/win/systlsdir.inc
index db53b77a94..32e0f28e36 100644
--- a/rtl/win/systlsdir.inc
+++ b/rtl/win/systlsdir.inc
@@ -32,8 +32,10 @@ Const
   DLL_PROCESS_DETACH = 0;
   DLL_THREAD_DETACH = 3;
 
+{$ifndef FPC_SECTION_THREADVARS}
 var
    TlsKey : PDWord = @TlsKeyVar;
+{$endif !FPC_SECTION_THREADVARS}
 
 type
   TTlsDirectory=packed record
@@ -44,11 +46,13 @@ type
   end;
 
 
+{$ifndef FPC_SECTION_THREADVARS}
 function TlsGetValue(dwTlsIndex : DWord) : pointer; stdcall;
   external 'kernel32' name 'TlsGetValue';
 
 procedure InitSystemThreads; external name '_FPC_InitSystemThreads';
 procedure SysAllocateThreadVars; external name '_FPC_SysAllocateThreadVars';
+{$endif !FPC_SECTION_THREADVARS}
 procedure InitHeap; external name '_FPC_InitHeap';
 
 {$ifdef FPC_HAS_INDIRECT_ENTRY_INFORMATION}
@@ -58,10 +62,9 @@ procedure SetupEntryInformation(constref info: TEntryInformation); external name
 {$endif FPC_INSSIDE_SYSINIT}
 
 procedure Exec_Tls_callback(Handle : pointer; reason : Dword; Reserved : pointer);
-  stdcall; [public,alias:'_FPC_Tls_Callback'];
+  stdcall;
   begin
-     if IsLibrary then
-       Exit;
+     Assert(not IsLibrary);
      case reason of
        { For executables, DLL_PROCESS_ATTACH is called *before* the entry point,
          and DLL_PROCESS_DETACH is called *after* RTL shuts down and calls ExitProcess.
@@ -83,12 +86,19 @@ procedure Exec_Tls_callback(Handle : pointer; reason : Dword; Reserved : pointer
              the threadvar handling can be correctly initialized }
            SetupEntryInformation(SysInitEntryInformation);
            {$endif FPC_HAS_INDIRECT_ENTRY_INFORMATION}
+           { the DLL_PROCESS_ATTACH initialization for the heap is required, because
+             if the DLL_PROCESS_ATTACH entry of a used DLL creates a thread, DLL_THREAD_ATTACH
+             of the executable will be called so that the RTL for the thread can be initialized
+             and that needs the code in DLL_PROCESS_ATTACH to be run first }
            InitHeap;
+         {$ifndef FPC_SECTION_THREADVARS}
            InitSystemThreads;
+         {$endif !FPC_SECTION_THREADVARS}
          end;
 
        DLL_THREAD_ATTACH :
          begin
+         {$ifndef FPC_SECTION_THREADVARS}
          {  !!! SysInitMultithreading must NOT be called here. Windows guarantees that
             the main thread invokes PROCESS_ATTACH, not THREAD_ATTACH. So this always
             executes in non-main thread. SysInitMultithreading() here will cause
@@ -100,6 +110,7 @@ procedure Exec_Tls_callback(Handle : pointer; reason : Dword; Reserved : pointer
          }
            { Allocate Threadvars  }
            SysAllocateThreadVars;
+         {$endif !FPC_SECTION_THREADVARS}
 
            { NS : no idea what is correct to pass here - pass dummy value for now }
            { passing a dummy is ok, the correct value is read from the coff header of SysInstance (FK) }
@@ -107,7 +118,11 @@ procedure Exec_Tls_callback(Handle : pointer; reason : Dword; Reserved : pointer
          end;
        DLL_THREAD_DETACH :
          begin
+         {$ifdef FPC_SECTION_THREADVARS}
+           if ThreadID<>0 then
+         {$else !FPC_SECTION_THREADVARS}
            if TlsGetValue(TLSKey^)<>nil then
+         {$endif !FPC_SECTION_THREADVARS}
              DoneThread; { Assume everything is idempotent there }
          end;
      end;
diff --git a/rtl/win/syswin.inc b/rtl/win/syswin.inc
index 63ee752569..37fb574f33 100644
--- a/rtl/win/syswin.inc
+++ b/rtl/win/syswin.inc
@@ -408,10 +408,12 @@ function Dll_entry{$ifdef FPC_HAS_INDIRECT_ENTRY_INFORMATION}(constref info : TE
          end;
        DLL_THREAD_ATTACH :
          begin
+         {$ifndef FPC_SECTION_THREADVARS}
            { SysInitMultithreading must not be called here,
              see comments in exec_tls_callback below }
            { Allocate Threadvars  }
            SysAllocateThreadVars;
+         {$endif !FPC_SECTION_THREADVARS}
 
            { NS : no idea what is correct to pass here - pass dummy value for now }
            { passing a dummy is ok, the correct value is read from the coff header of SysInstance (FK) }
@@ -425,7 +427,11 @@ function Dll_entry{$ifdef FPC_HAS_INDIRECT_ENTRY_INFORMATION}(constref info : TE
            if assigned(Dll_Thread_Detach_Hook) then
              Dll_Thread_Detach_Hook(DllParam);
            { Release Threadvars }
+         {$ifdef FPC_SECTION_THREADVARS}
+           if ThreadID<>0 then
+         {$else !FPC_SECTION_THREADVARS}
            if TlsGetValue(TLSKey^)<>nil then
+         {$endif !FPC_SECTION_THREADVARS}
              DoneThread; { Assume everything is idempotent there }
          end;
        DLL_PROCESS_DETACH :
@@ -439,9 +445,11 @@ function Dll_entry{$ifdef FPC_HAS_INDIRECT_ENTRY_INFORMATION}(constref info : TE
                InternalExit;
              end;
 
+         {$ifndef FPC_SECTION_THREADVARS}
            SysReleaseThreadVars;
            { Free TLS resources used by ThreadVars }
            SysFiniMultiThreading;
+         {$endif !FPC_SECTION_THREADVARS}
            MainThreadIDWin32:=0;
          end;
      end;
diff --git a/rtl/win32/sysinit.inc b/rtl/win32/sysinit.inc
index 0eea501af0..c8f7b6f042 100644
--- a/rtl/win32/sysinit.inc
+++ b/rtl/win32/sysinit.inc
@@ -16,10 +16,14 @@
 
    var
       SysInstance : LongInt;
+ {$ifndef FPC_SECTION_THREADVARS}
       TlsKeyVar: DWord = $ffffffff;
+ {$endif !FPC_SECTION_THREADVARS}
 
       InitFinalTable : record end; external name 'INITFINAL';
+ {$ifndef FPC_SECTION_THREADVARS}
       ThreadvarTablesTable : record end; external name 'FPC_THREADVARTABLES';
+ {$endif !FPC_SECTION_THREADVARS}
       WideInitTables : record end; external name 'FPC_WIDEINITTABLES';
  {$ifdef FPC_HAS_RESSTRINITS}
       ResStrInitTables : record end; external name 'FPC_RESSTRINITTABLES';
@@ -53,7 +57,11 @@
       STD_INPUT_HANDLE = dword(-10);
       SysInitEntryInformation : TEntryInformation = (
         InitFinalTable : @InitFinalTable;
+{$ifdef FPC_SECTION_THREADVARS}
+        ThreadvarTablesTable : nil;
+{$else FPC_SECTION_THREADVARS}
         ThreadvarTablesTable : @ThreadvarTablesTable;
+{$endif FPC_SECTION_THREADVARS}
         ResourceStringTables : @ResourceStringTables;
 {$ifdef FPC_HAS_RESSTRINITS}
         ResStrInitTables : @ResStrInitTables;
@@ -65,7 +73,9 @@
         valgrind_used : false;
         OS : (
           asm_exit : @asm_exit;
+{$ifndef FPC_SECTION_THREADVARS}
           TlsKeyAddr : @TlsKeyVar;
+{$endif !FPC_SECTION_THREADVARS}
           SysInstance : @SysInstance;
           WideInitTables: @WideInitTables;
           );
diff --git a/rtl/win32/system.pp b/rtl/win32/system.pp
index 4ffa0b65ff..4cf69de94d 100644
--- a/rtl/win32/system.pp
+++ b/rtl/win32/system.pp
@@ -78,7 +78,9 @@ end;
 
 procedure OsSetupEntryInformation(constref info: TEntryInformation);
 begin
+{$ifndef FPC_SECTION_THREADVARS}
   TlsKey := info.OS.TlsKeyAddr;
+{$endif !FPC_SECTION_THREADVARS}
   FPCSysInstance := info.OS.SysInstance;
   WStrInitTablesTable := info.OS.WideInitTables;
 end;
@@ -498,41 +500,6 @@ begin
 end;
 {$endif Set_i386_Exception_handler}
 
-{$ifdef FPC_SECTION_THREADVARS}
-function fpc_tls_add(addr: pointer): pointer; assembler; nostackframe;
-  [public,alias: 'FPC_TLS_ADD']; compilerproc;
-  asm
-      sub   $tls_data_start,%eax
-      cmpb  $0,IsLibrary
-      mov   _tls_index,%ecx
-      jnz   .L1
-      mov   %fs:(0x2c),%edx
-      add   (%edx,%ecx,4),%eax
-      ret
-.L1:
-      push  %ebx
-      mov   %eax,%ebx
-      call  GetLastError
-      push  %eax                      { save LastError }
-      push  _tls_index
-      call  TlsGetValue
-      test  %eax,%eax
-      jnz   .L2
-      { This can happen when a thread existed before DLL was loaded,
-        or if DisableThreadLibraryCalls was called. }
-      call  SysAllocateThreadVars
-      mov   $0x1000000,%eax
-      call  InitThread
-      push  _tls_index
-      call  TlsGetValue
-.L2:
-      add   %eax,%ebx
-      call  SetLastError              { restore (value is on stack) }
-      mov   %ebx,%eax
-      pop   %ebx
-  end;
-{$endif FPC_SECTION_THREADVARS}
-
 function CheckInitialStkLen(stklen : SizeUInt) : SizeUInt;
   type
     tdosheader = packed record
diff --git a/rtl/win64/sysinit.pp b/rtl/win64/sysinit.pp
index 5d396beb8f..f658cccb2b 100644
--- a/rtl/win64/sysinit.pp
+++ b/rtl/win64/sysinit.pp
@@ -21,10 +21,14 @@ unit sysinit;
 
    var
       SysInstance : QWord;
+{$ifndef FPC_SECTION_THREADVARS}
       TlsKeyVar: DWord = $ffffffff;
+{$endif !FPC_SECTION_THREADVARS}
 
       InitFinalTable : record end; external name 'INITFINAL';
+{$ifndef FPC_SECTION_THREADVARS}
       ThreadvarTablesTable : record end; external name 'FPC_THREADVARTABLES';
+{$endif !FPC_SECTION_THREADVARS}
       WideInitTables : record end; external name 'FPC_WIDEINITTABLES';
 {$ifdef FPC_HAS_RESSTRINITS}
       ResStrInitTables : record end; external name 'FPC_RESSTRINITTABLES';
@@ -58,7 +62,11 @@ unit sysinit;
       STD_INPUT_HANDLE = dword(-10);
       SysInitEntryInformation : TEntryInformation = (
         InitFinalTable : @InitFinalTable;
+{$ifdef FPC_SECTION_THREADVARS}
+        ThreadvarTablesTable : nil;
+{$else FPC_SECTION_THREADVARS}
         ThreadvarTablesTable : @ThreadvarTablesTable;
+{$endif FPC_SECTION_THREADVARS}
         ResourceStringTables : @ResourceStringTables;
 {$ifdef FPC_HAS_RESSTRINITS}
         ResStrInitTables : @ResStrInitTables;
@@ -69,7 +77,9 @@ unit sysinit;
         PascalMain : @PascalMain;
         valgrind_used : false;
         OS : (
+{$ifndef FPC_SECTION_THREADVARS}
           TlsKeyAddr : @TlsKeyVar;
+{$endif !FPC_SECTION_THREADVARS}
           SysInstance : @SysInstance;
           WideInitTables: @WideInitTables;
           );
@@ -124,6 +134,9 @@ unit sysinit;
       sysinstance:=_hinstance;
       dllreason:=_dllreason;
       dllparam:=PtrInt(_dllparam);
+{$if defined(FPC_USE_TLS_DIRECTORY) and defined(FPC_SECTION_THREADVARS)}
+      LinkIn(@tlsdir,@tls_callback_end,nil);
+{$endif}
       SetupEntryInformation;
       DLL_Entry(SysInitEntryInformation);
     end;
@@ -135,6 +148,9 @@ unit sysinit;
       sysinstance:=_hinstance;
       dllreason:=_dllreason;
       dllparam:=PtrInt(_dllparam);
+{$if defined(FPC_USE_TLS_DIRECTORY) and defined(FPC_SECTION_THREADVARS)}
+      LinkIn(@tlsdir,@tls_callback_end,nil);
+{$endif}
       SetupEntryInformation;
       DLL_Entry(SysInitEntryInformation);
     end;
diff --git a/rtl/win64/system.pp b/rtl/win64/system.pp
index 6cf7b174cd..b6fad06fff 100644
--- a/rtl/win64/system.pp
+++ b/rtl/win64/system.pp
@@ -122,7 +122,9 @@ procedure PascalMain;external name 'PASCALMAIN';
 
 procedure OsSetupEntryInformation(constref info: TEntryInformation);
 begin
+{$ifndef FPC_SECTION_THREADVARS}
   TlsKey := info.OS.TlsKeyAddr;
+{$endif !FPC_SECTION_THREADVARS}
   FPCSysInstance := info.OS.SysInstance;
   WStrInitTablesTable := info.OS.WideInitTables;
 end;
@@ -504,47 +506,6 @@ begin
 end;
 {$endif VER3_0}
 
-{$ifdef FPC_SECTION_THREADVARS}
-function fpc_tls_add(addr: pointer): pointer; assembler; nostackframe;
-  [public,alias: 'FPC_TLS_ADD']; compilerproc;
-  asm
-      sub   $56,%rsp                  { 32 spill area + 16 local vars + 8 misalignment }
-  .seh_stackalloc 56
-  .seh_endprologue
-      lea   tls_data_start(%rip),%rax
-      sub   %rax,%rcx
-      cmpb  $0,IsLibrary(%rip)
-      mov   _tls_index(%rip),%eax
-      jnz   .L1
-      mov   %gs:(88),%rdx
-      add   (%rdx,%rax,8),%rcx
-      mov   %rcx,%rax
-      jmp   .L3
-.L1:
-      mov   %rcx,32(%rsp)
-      call  GetLastError
-      mov   %rax,40(%rsp)             { save LastError }
-      mov   _tls_index(%rip),%ecx
-      call  TlsGetValue
-      test  %rax,%rax
-      jnz   .L2
-      { This can happen when a thread existed before DLL was loaded,
-        or if DisableThreadLibraryCalls was called. }
-      call  SysAllocateThreadVars
-      mov   $0x1000000,%rcx
-      call  InitThread
-      mov   _tls_index(%rip),%ecx
-      call  TlsGetValue
-.L2:
-      add   %rax,32(%rsp)
-      mov   40(%rsp),%rcx
-      call  SetLastError
-      mov   32(%rsp),%rax
-.L3:
-      add   $56,%rsp
-  end;
-{$endif FPC_SECTION_THREADVARS}
-
 function CheckInitialStkLen(stklen : SizeUInt) : SizeUInt;
   type
     tdosheader = packed record
-- 
2.26.2.windows.1

---
 rtl/win/systlsdir.inc | 3 +--
 rtl/win64/sysinit.pp  | 4 ++--
 2 files changed, 3 insertions(+), 4 deletions(-)

diff --git a/rtl/win/systlsdir.inc b/rtl/win/systlsdir.inc
index d0b48a89a5..32e0f28e36 100644
--- a/rtl/win/systlsdir.inc
+++ b/rtl/win/systlsdir.inc
@@ -64,8 +64,7 @@ procedure SetupEntryInformation(constref info: TEntryInformation); external name
 procedure Exec_Tls_callback(Handle : pointer; reason : Dword; Reserved : pointer);
   stdcall;
   begin
-     if IsLibrary then
-       Exit;
+     Assert(not IsLibrary);
      case reason of
        { For executables, DLL_PROCESS_ATTACH is called *before* the entry point,
          and DLL_PROCESS_DETACH is called *after* RTL shuts down and calls ExitProcess.
diff --git a/rtl/win64/sysinit.pp b/rtl/win64/sysinit.pp
index 9d740e19ec..f658cccb2b 100644
--- a/rtl/win64/sysinit.pp
+++ b/rtl/win64/sysinit.pp
@@ -135,7 +135,7 @@ unit sysinit;
       dllreason:=_dllreason;
       dllparam:=PtrInt(_dllparam);
 {$if defined(FPC_USE_TLS_DIRECTORY) and defined(FPC_SECTION_THREADVARS)}
-      LinkIn(@tlsdir,@tls_callback_end,@tls_callback);
+      LinkIn(@tlsdir,@tls_callback_end,nil);
 {$endif}
       SetupEntryInformation;
       DLL_Entry(SysInitEntryInformation);
@@ -149,7 +149,7 @@ unit sysinit;
       dllreason:=_dllreason;
       dllparam:=PtrInt(_dllparam);
 {$if defined(FPC_USE_TLS_DIRECTORY) and defined(FPC_SECTION_THREADVARS)}
-      LinkIn(@tlsdir,@tls_callback_end,@tls_callback);
+      LinkIn(@tlsdir,@tls_callback_end,nil);
 {$endif}
       SetupEntryInformation;
       DLL_Entry(SysInitEntryInformation);
-- 
2.26.2.windows.1

---
 compiler/systems/t_win.pas | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/compiler/systems/t_win.pas b/compiler/systems/t_win.pas
index 887787f8e9..583b71088e 100644
--- a/compiler/systems/t_win.pas
+++ b/compiler/systems/t_win.pas
@@ -1343,7 +1343,7 @@ implementation
             Add('    ___crt_xi_start__ = . ;');
             Add('    *(SORT(.CRT$XI*))  /* C++ initialization */');
             Add('    ___crt_xi_end__ = . ;');
-            Add('    ___crt_xl_start__ = . ;');
+            Add('    PROVIDE (___crt_xl_start__ = .);');
             Add('    *(SORT(.CRT$XL*))  /* TLS callbacks */');
             Add('    /* ___crt_xl_end__ is defined in the TLS Directory support code */');
             Add('    PROVIDE (___crt_xl_end__ = .);');
-- 
2.26.2.windows.1

---
 rtl/inc/thread.inc  | 4 ++++
 rtl/win64/system.pp | 4 ++++
 2 files changed, 8 insertions(+)

diff --git a/rtl/inc/thread.inc b/rtl/inc/thread.inc
index ce7ec796b4..07e20bb145 100644
--- a/rtl/inc/thread.inc
+++ b/rtl/inc/thread.inc
@@ -52,16 +52,20 @@ Var
           widestringmanager.ThreadInitProc;
 {$endif FPC_HAS_FEATURE_WIDESTRINGS}
 {$ifdef FPC_HAS_FEATURE_EXCEPTIONS}
+{$ifndef FPC_SECTION_THREADVARS}
         { ExceptAddrStack and ExceptObjectStack are threadvars       }
         { so every thread has its on exception handling capabilities }
         SysInitExceptions;
+{$endif !FPC_SECTION_THREADVARS}
 {$endif FPC_HAS_FEATURE_EXCEPTIONS}
 {$ifdef FPC_HAS_FEATURE_CONSOLEIO}
 {$ifndef EMBEDDED}
         { Open all stdio fds again }
         SysInitStdio;
+{$ifndef FPC_SECTION_THREADVARS}
         InOutRes:=0;
         // ErrNo:=0;
+{$endif !FPC_SECTION_THREADVARS}
 {$endif EMBEDDED}
 {$endif FPC_HAS_FEATURE_CONSOLEIO}
 {$ifdef FPC_HAS_FEATURE_STACKCHECK}
diff --git a/rtl/win64/system.pp b/rtl/win64/system.pp
index b6fad06fff..a5771e281e 100644
--- a/rtl/win64/system.pp
+++ b/rtl/win64/system.pp
@@ -593,15 +593,19 @@ initialization
     InitHeap;
     InitSystemThreads;
   end;
+{$ifndef FPC_SECTION_THREADVARS}
   SysInitExceptions;
+{$endif !FPC_SECTION_THREADVARS}
   initunicodestringmanager;
   InitWin32Widestrings;
   SysInitStdIO;
   { Arguments }
   setup_arguments;
   InitSystemDynLibs;
+{$ifndef FPC_SECTION_THREADVARS}
   { Reset IO Error }
   InOutRes:=0;
+{$endif !FPC_SECTION_THREADVARS}
   ProcessID := GetCurrentProcessID;
   DispCallByIDProc:=@DoDispCallByIDError;
 
-- 
2.26.2.windows.1

---
 compiler/globals.pas    | 2 ++
 compiler/options.pas    | 3 +++
 compiler/pmodules.pas   | 6 ++++++
 compiler/x86/nx86ld.pas | 4 ++--
 4 files changed, 13 insertions(+), 2 deletions(-)

diff --git a/compiler/globals.pas b/compiler/globals.pas
index c3a1f781a0..78d8fb26eb 100644
--- a/compiler/globals.pas
+++ b/compiler/globals.pas
@@ -165,6 +165,7 @@ interface
          disabledircache : boolean;
 
          tlsmodel : ttlsmodel;
+         tlsmodel_auto : boolean;
 
 {$if defined(i8086)}
          x86memorymodel  : tx86memorymodel;
@@ -585,6 +586,7 @@ interface
         disabledircache : false;
 
         tlsmodel : tlsm_none;
+        tlsmodel_auto : false;
 {$if defined(i8086)}
         x86memorymodel : mm_small;
 {$endif defined(i8086)}
diff --git a/compiler/options.pas b/compiler/options.pas
index b77220e47a..65fa7611e4 100644
--- a/compiler/options.pas
+++ b/compiler/options.pas
@@ -4250,8 +4250,11 @@ begin
     begin
       if cs_create_pic in init_settings.moduleswitches then
         init_settings.tlsmodel:=tlsm_global_dynamic
+      else if target_info.system in [system_x86_64_win64] then
+        init_settings.tlsmodel:=tlsm_initial_exec
       else
         init_settings.tlsmodel:=tlsm_local_exec;
+      init_settings.tlsmodel_auto:=true;
     end;
 
   { set Mac OS X version default macros if not specified explicitly }
diff --git a/compiler/pmodules.pas b/compiler/pmodules.pas
index bf66024f02..9378377127 100644
--- a/compiler/pmodules.pas
+++ b/compiler/pmodules.pas
@@ -1962,6 +1962,12 @@ type
          { make the compiler happy and avoid an uninitialized variable warning on Setlength(sc,length(sc)+1); }
          sc:=nil;
 
+         if not IsLibrary and init_settings.tlsmodel_auto and (init_settings.tlsmodel=tlsm_initial_exec) then
+           begin
+             init_settings.tlsmodel:=tlsm_local_exec;
+             current_settings.tlsmodel:=tlsm_local_exec;
+           end;
+
          { DLL defaults to create reloc info }
          if islibrary or (target_info.system in [system_aarch64_win64]) then
            begin
diff --git a/compiler/x86/nx86ld.pas b/compiler/x86/nx86ld.pas
index 61c82755d4..d9e07bad0e 100644
--- a/compiler/x86/nx86ld.pas
+++ b/compiler/x86/nx86ld.pas
@@ -38,7 +38,7 @@ interface
 implementation
 
     uses
-      globals,fmodule,
+      globals,
       cutils,verbose,systems,
       aasmbase,aasmtai,aasmdata,aasmcpu,
       cgutils,cgobj,hlcgobj,
@@ -136,7 +136,7 @@ implementation
                   current_asmdata.CurrAsmList.concat(taicpu.op_ref_reg(A_MOV,S_Q,href,hregister));         // movq %gs:88, %rcx
 
                   reference_reset_base(href,hregister,0,ctempposinvalid,sizeof(AInt),[]);
-                  if (main_module=nil) or main_module.IsLibrary or main_module.is_unit then
+                  if current_settings.tlsmodel<>tlsm_local_exec then
                     begin // _tls_index <> 0
                       reference_reset_symbol(href_index,current_asmdata.RefAsmSymbol('_tls_index',AT_DATA),0,sizeof(dword),[]);
                       reg_index:=cg.getintregister(current_asmdata.CurrAsmList,OS_INT);
-- 
2.26.2.windows.1

---
 compiler/aggas.pas         |  4 +-
 compiler/cgbase.pas        |  1 +
 compiler/dbgdwarf.pas      |  3 +-
 compiler/globals.pas       |  2 +
 compiler/nld.pas           |  4 +-
 compiler/nutils.pas        |  9 +++--
 compiler/ogcoff.pas        | 35 +++++++----------
 compiler/options.pas       |  3 ++
 compiler/pmodules.pas      |  6 +++
 compiler/systems/i_win.pas |  3 ++
 compiler/systems/t_win.pas | 27 ++++++++++++-
 compiler/x86/aasmcpu.pas   |  6 +++
 compiler/x86/agx86att.pas  |  9 +++++
 compiler/x86/nx86ld.pas    | 77 +++++++++++++++++++++++---------------
 rtl/inc/heaptrc.pp         |  4 ++
 rtl/inc/thread.inc         |  4 ++
 rtl/linux/si_impl.inc      |  2 +
 rtl/win/sysosh.inc         |  5 +--
 rtl/win/systhrd.inc        | 27 ++++++++++---
 rtl/win/systlsdir.inc      | 21 +++++++++--
 rtl/win/syswin.inc         |  8 ++++
 rtl/win32/sysinit.inc      | 10 +++++
 rtl/win32/system.pp        | 37 +-----------------
 rtl/win64/sysinit.pp       | 16 ++++++++
 rtl/win64/system.pp        | 47 +++--------------------
 25 files changed, 222 insertions(+), 148 deletions(-)

diff --git a/compiler/aggas.pas b/compiler/aggas.pas
index 41d20c5129..f9b69f605b 100644
--- a/compiler/aggas.pas
+++ b/compiler/aggas.pas
@@ -533,7 +533,9 @@ implementation
              { according to the GNU AS guide AS for COFF does not support the
                progbits }
              writer.AsmWrite('.section ');
-             usesectionflags:=true;
+             { .tls section requires special processing }
+             if not((atype=sec_threadvar) and (tf_section_threadvars in target_info.flags)) then
+               usesectionflags:=true;
            end;
          system_powerpc_darwin,
          system_i386_darwin,
diff --git a/compiler/cgbase.pas b/compiler/cgbase.pas
index 6dca71ba40..0e862ca739 100644
--- a/compiler/cgbase.pas
+++ b/compiler/cgbase.pas
@@ -140,6 +140,7 @@ interface
 {$ifdef x86_64}
           ,addr_tpoff
           ,addr_tlsgd
+          ,addr_secrel
 {$endif x86_64}
          );
 
diff --git a/compiler/dbgdwarf.pas b/compiler/dbgdwarf.pas
index 5b75e257b5..d81436ed2a 100644
--- a/compiler/dbgdwarf.pas
+++ b/compiler/dbgdwarf.pas
@@ -2658,7 +2658,8 @@ implementation
                   begin
                     if vo_is_thread_var in sym.varoptions then
                       begin
-                        if tf_section_threadvars in target_info.flags then
+                        if (tf_section_threadvars in target_info.flags) and
+                           (not (target_info.system in [system_x86_64_win64])) then // TODO : dwarf for threadvars on win64
                           begin
                             case sizeof(puint) of
                               2:
diff --git a/compiler/globals.pas b/compiler/globals.pas
index c3a1f781a0..78d8fb26eb 100644
--- a/compiler/globals.pas
+++ b/compiler/globals.pas
@@ -165,6 +165,7 @@ interface
          disabledircache : boolean;
 
          tlsmodel : ttlsmodel;
+         tlsmodel_auto : boolean;
 
 {$if defined(i8086)}
          x86memorymodel  : tx86memorymodel;
@@ -585,6 +586,7 @@ interface
         disabledircache : false;
 
         tlsmodel : tlsm_none;
+        tlsmodel_auto : false;
 {$if defined(i8086)}
         x86memorymodel : mm_small;
 {$endif defined(i8086)}
diff --git a/compiler/nld.pas b/compiler/nld.pas
index e2bccbc080..0453f162e8 100644
--- a/compiler/nld.pas
+++ b/compiler/nld.pas
@@ -450,7 +450,9 @@ implementation
                 { call to get address of threadvar }
                 if (vo_is_thread_var in tabstractvarsym(symtableentry).varoptions) then
                   begin
-                    include(current_procinfo.flags,pi_do_call);
+                    if (not (tf_section_threadvars in target_info.flags)) or
+                       (current_settings.tlsmodel in [tlsm_local_dynamic,tlsm_global_dynamic]) then
+                      include(current_procinfo.flags,pi_do_call);
                     include(current_procinfo.flags,pi_uses_threadvar);
                   end;
               end;
diff --git a/compiler/nutils.pas b/compiler/nutils.pas
index d586fc222d..b5c0de1937 100644
--- a/compiler/nutils.pas
+++ b/compiler/nutils.pas
@@ -191,7 +191,7 @@ interface
 implementation
 
     uses
-      cutils,verbose,globals,
+      cutils,verbose,globals,systems,
       symconst,symdef,
       defcmp,defutil,
       ncon,ncnv,nld,nflw,nset,ncal,nadd,nmem,ninl,
@@ -799,10 +799,13 @@ implementation
                 begin
                   if assigned(tloadnode(p).left) then
                     inc(result,node_complexity(tloadnode(p).left));
-                  { threadvars need a helper call }
+                  { some threadvars models need a helper call }
                   if (tloadnode(p).symtableentry.typ=staticvarsym) and
                      (vo_is_thread_var in tstaticvarsym(tloadnode(p).symtableentry).varoptions) then
-                    inc(result,5)
+                    if (tf_section_threadvars in target_info.flags) and (current_settings.tlsmodel in [tlsm_local_exec,tlsm_initial_exec]) then
+                      inc(result,3)
+                    else
+                      inc(result,50)
                   else if not((tloadnode(p).symtableentry.typ in [staticvarsym,localvarsym,paravarsym,fieldvarsym]) and
                     (tabstractvarsym(tloadnode(p).symtableentry).varregable in [vr_intreg,vr_mmreg,vr_fpureg])) then
                     inc(result);
diff --git a/compiler/ogcoff.pas b/compiler/ogcoff.pas
index 8c05b3792a..90bbca8b0d 100644
--- a/compiler/ogcoff.pas
+++ b/compiler/ogcoff.pas
@@ -1015,8 +1015,10 @@ const pemagic : array[0..3] of byte = (
                   end;
                 RELOC_SECREL32 :
                   begin
-                    { fixup address when the symbol was known in defined object }
-                    if (relocsec.objdata=objsec.objdata) then
+                    { fixup address when the symbol was known in defined object or
+                      when the symbol is a threadvar }
+                    if (relocsec.objdata=objsec.objdata) or
+                       assigned(objreloc.symbol) and (copy(objreloc.symbol.objsection.Name,1,4)='.tls') then
                       dec(address,relocsec.ExeSection.MemPos);
                     inc(address,relocval);
                   end;
@@ -2667,21 +2669,17 @@ const pemagic : array[0..3] of byte = (
 
         procedure UpdateTlsDataDir;
         var
-          {callbacksection : TExeSection;}
           tlsexesymbol: TExeSymbol;
           tlssymbol: TObjSymbol;
-          callbackexesymbol: TExeSymbol;
-          //callbacksymbol: TObjSymbol;
         begin
           { according to GNU ld,
             the callback routines should be placed into .CRT$XL*
             sections, and the thread local variables in .tls
             __tls_start__ and __tls_end__ symbols
-            should be used for the initialized part,
-            which we do not support yet. }
-          { For now, we only pass the address of the __tls_used
+            should be used for the initialized part. }
+          { We pass the address of the __tls_used
             asm symbol into PE_DATADIR_TLS with the correct
-            size of this table (different for win32/win64 }
+            size of this table (different for win32/win64) }
           tlsexesymbol:=texesymbol(ExeSymbolList.Find(
             target_info.Cprefix+'_tls_used'));
           if assigned(tlsexesymbol) then
@@ -2690,18 +2688,6 @@ const pemagic : array[0..3] of byte = (
               peoptheader.DataDirectory[PE_DATADIR_TLS].vaddr:=tlssymbol.address;
               { sizeof(TlsDirectory) is different on host and target when cross-compiling }
               peoptheader.DataDirectory[PE_DATADIR_TLS].size:=TLSDIR_SIZE;
-              if IsSharedLibrary then
-                begin
-                  { Here we should reset __FPC_tls_callbacks value to nil }
-                  callbackexesymbol:=texesymbol(ExeSymbolList.Find(
-                                        '__FPC_tls_callbacks'));
-                  if assigned (callbackexesymbol) then
-                    begin
-                      //callbacksymbol:=callbackexesymbol.ObjSymbol;
-
-                    end;
-                end;
-
            end;
         end;
 
@@ -2790,7 +2776,10 @@ const pemagic : array[0..3] of byte = (
               end
             else
               begin
-                peoptheader.MajorOperatingSystemVersion:=4;
+                if tf_section_threadvars in target_info.flags then
+                  peoptheader.MajorOperatingSystemVersion:=6 // >= Windows Vista
+                else
+                  peoptheader.MajorOperatingSystemVersion:=4;
                 peoptheader.MinorOperatingSystemVersion:=0;
               end;
             if SetPEUserVersionSetExplicitely then
@@ -2812,6 +2801,8 @@ const pemagic : array[0..3] of byte = (
               begin
                 if target_info.system in systems_wince then
                   peoptheader.MajorSubsystemVersion:=3
+                else if tf_section_threadvars in target_info.flags then
+                  peoptheader.MajorSubsystemVersion:=6 // >= Windows Vista
                 else
                   peoptheader.MajorSubsystemVersion:=4;
                 peoptheader.MinorSubsystemVersion:=0;
diff --git a/compiler/options.pas b/compiler/options.pas
index b77220e47a..65fa7611e4 100644
--- a/compiler/options.pas
+++ b/compiler/options.pas
@@ -4250,8 +4250,11 @@ begin
     begin
       if cs_create_pic in init_settings.moduleswitches then
         init_settings.tlsmodel:=tlsm_global_dynamic
+      else if target_info.system in [system_x86_64_win64] then
+        init_settings.tlsmodel:=tlsm_initial_exec
       else
         init_settings.tlsmodel:=tlsm_local_exec;
+      init_settings.tlsmodel_auto:=true;
     end;
 
   { set Mac OS X version default macros if not specified explicitly }
diff --git a/compiler/pmodules.pas b/compiler/pmodules.pas
index bf66024f02..9378377127 100644
--- a/compiler/pmodules.pas
+++ b/compiler/pmodules.pas
@@ -1962,6 +1962,12 @@ type
          { make the compiler happy and avoid an uninitialized variable warning on Setlength(sc,length(sc)+1); }
          sc:=nil;
 
+         if not IsLibrary and init_settings.tlsmodel_auto and (init_settings.tlsmodel=tlsm_initial_exec) then
+           begin
+             init_settings.tlsmodel:=tlsm_local_exec;
+             current_settings.tlsmodel:=tlsm_local_exec;
+           end;
+
          { DLL defaults to create reloc info }
          if islibrary or (target_info.system in [system_aarch64_win64]) then
            begin
diff --git a/compiler/systems/i_win.pas b/compiler/systems/i_win.pas
index d4e8f842d5..d1ffb4e57c 100644
--- a/compiler/systems/i_win.pas
+++ b/compiler/systems/i_win.pas
@@ -107,6 +107,9 @@ unit i_win;
             shortname    : 'Win64';
             flags        : [tf_files_case_aware,tf_has_dllscanner,
                             tf_smartlink_sections,
+{$if not defined(DISABLE_TLS_DIRECTORY) and not defined(SUPPORT_WINXP)}
+                            tf_section_threadvars,
+{$endif}
                             tf_winlikewidestring,tf_no_pic_supported,
                             tf_dwarf_only_local_labels,
                             tf_no_generic_stackcheck,tf_has_winlike_resources,
diff --git a/compiler/systems/t_win.pas b/compiler/systems/t_win.pas
index 887787f8e9..b96f820a3c 100644
--- a/compiler/systems/t_win.pas
+++ b/compiler/systems/t_win.pas
@@ -1711,11 +1711,36 @@ implementation
             else
               ;
           end;
-        if dllversion<>'' then
+        if SetPEUserVersionSetExplicitely then
+          begin
+            peoptheader.MajorImageVersion:=peuserversionmajor;
+            peoptheader.MinorImageVersion:=peuserversionminor;
+          end
+        else if dllversion<>'' then
           begin
            peoptheader.MajorImageVersion:=dllmajor;
            peoptheader.MinorImageVersion:=dllminor;
           end;
+        if SetPEOSVersionSetExplicitely then
+          begin
+            peoptheader.MajorOperatingSystemVersion:=peosversionmajor;
+            peoptheader.MinorOperatingSystemVersion:=peosversionminor;
+          end
+        else if tf_section_threadvars in target_info.flags then
+          begin
+            peoptheader.MajorOperatingSystemVersion:=6; // >= Windows Vista
+            peoptheader.MinorOperatingSystemVersion:=0;
+          end;
+        if SetPESubSysVersionSetExplicitely then
+          begin
+            peoptheader.MajorSubsystemVersion:=pesubsysversionmajor;
+            peoptheader.MinorSubsystemVersion:=pesubsysversionminor;
+          end
+        else if tf_section_threadvars in target_info.flags then
+          begin
+            peoptheader.MajorSubsystemVersion:=6; // >= Windows Vista
+            peoptheader.MinorSubsystemVersion:=0;
+          end;
         { reset timestamp }
         peheader.time:=0;
         { write header back, skip pe magic }
diff --git a/compiler/x86/aasmcpu.pas b/compiler/x86/aasmcpu.pas
index f712af7196..115474e6d1 100644
--- a/compiler/x86/aasmcpu.pas
+++ b/compiler/x86/aasmcpu.pas
@@ -3617,6 +3617,12 @@ implementation
                       currabsreloc:=RELOC_TLSGD;
                       currabsreloc32:=RELOC_TLSGD;
                     end
+                  else if oper[opidx]^.ref^.refaddr=addr_secrel then
+                    begin
+                      currrelreloc:=RELOC_SECREL32;
+                      currabsreloc:=RELOC_SECREL32;
+                      currabsreloc32:=RELOC_SECREL32;
+                    end
                   else
 {$endif x86_64}
                     begin
diff --git a/compiler/x86/agx86att.pas b/compiler/x86/agx86att.pas
index 657f073163..bf69e08bd5 100644
--- a/compiler/x86/agx86att.pas
+++ b/compiler/x86/agx86att.pas
@@ -190,6 +190,8 @@ interface
                owner.writer.AsmWrite('@tpoff');
              addr_tlsgd:
                owner.writer.AsmWrite('@tlsgd');
+             addr_secrel:
+               owner.writer.AsmWrite('@secrel32');
 {$endif x86_64}
              else
                ;
@@ -259,6 +261,13 @@ interface
               {$ifdef x86_64},addr_tpoff,addr_tlsgd{$endif x86_64}
               ] then
               WriteReference(o.ref^)
+        {$ifdef x86_64}
+            else if o.ref^.refaddr=addr_secrel then
+              begin
+                owner.writer.AsmWrite('$');
+                WriteReference(o.ref^);
+              end
+        {$endif x86_64}
             else
               begin
                 owner.writer.AsmWrite('$');
diff --git a/compiler/x86/nx86ld.pas b/compiler/x86/nx86ld.pas
index 774addd1ca..d9e07bad0e 100644
--- a/compiler/x86/nx86ld.pas
+++ b/compiler/x86/nx86ld.pas
@@ -41,7 +41,7 @@ implementation
       globals,
       cutils,verbose,systems,
       aasmbase,aasmtai,aasmdata,aasmcpu,
-      cgutils,cgobj,
+      cgutils,cgobj,hlcgobj,
       symconst,symdef,symtable,
       cgbase,cpubase,parabase,paramgr,
       procinfo;
@@ -54,37 +54,10 @@ implementation
       var
         paraloc1 : tcgpara;
         pd: tprocdef;
-        href: treference;
-        hregister : tregister;
-        handled: boolean;
+        href, href_index: treference;
+        hregister, reg_index, reg_offset, reg_start : tregister;
       begin
-        handled:=false;
-        if (tf_section_threadvars in target_info.flags) then
-          begin
-            if target_info.system in [system_i386_win32,system_x86_64_win64] then
-              begin
-                paraloc1.init;
-                pd:=search_system_proc('fpc_tls_add');
-                paramanager.getcgtempparaloc(current_asmdata.CurrAsmList,pd,1,paraloc1);
-                if not(vo_is_weak_external in gvs.varoptions) then
-                  reference_reset_symbol(href,current_asmdata.RefAsmSymbol(gvs.mangledname,AT_DATA,use_indirect_symbol(gvs)),0,sizeof(pint),[])
-                else
-                  reference_reset_symbol(href,current_asmdata.WeakRefAsmSymbol(gvs.mangledname,AT_DATA),0,sizeof(pint),[]);
-                cg.a_loadaddr_ref_cgpara(current_asmdata.CurrAsmList,href,paraloc1);
-                paramanager.freecgpara(current_asmdata.CurrAsmList,paraloc1);
-                paraloc1.done;
-
-                cg.g_call(current_asmdata.CurrAsmList,'FPC_TLS_ADD');
-                cg.ungetcpuregister(current_asmdata.CurrAsmList,NR_FUNCTION_RESULT_REG);
-                hregister:=cg.getaddressregister(current_asmdata.CurrAsmList);
-                cg.a_load_reg_reg(current_asmdata.CurrAsmList,OS_ADDR,OS_ADDR,NR_FUNCTION_RESULT_REG,hregister);
-                location.reference.base:=hregister;
-                handled:=true;
-              end;
-          end;
-
-        if not handled then
-          inherited;
+        inherited;
 
         if (tf_section_threadvars in target_info.flags) then
           begin
@@ -156,6 +129,48 @@ implementation
                       Internalerror(2019012002);
                   end;
                 end;
+              system_x86_64_win64:
+                begin
+                  reference_reset(href,sizeof(AInt),[]); href.segment:=NR_GS; href.offset:=88;
+                  hregister:=cg.getaddressregister(current_asmdata.CurrAsmList);
+                  current_asmdata.CurrAsmList.concat(taicpu.op_ref_reg(A_MOV,S_Q,href,hregister));         // movq %gs:88, %rcx
+
+                  reference_reset_base(href,hregister,0,ctempposinvalid,sizeof(AInt),[]);
+                  if current_settings.tlsmodel<>tlsm_local_exec then
+                    begin // _tls_index <> 0
+                      reference_reset_symbol(href_index,current_asmdata.RefAsmSymbol('_tls_index',AT_DATA),0,sizeof(dword),[]);
+                      reg_index:=cg.getintregister(current_asmdata.CurrAsmList,OS_INT);
+                      hlcg.a_load_ref_reg(current_asmdata.CurrAsmList,u32inttype,u64inttype,
+                        href_index,reg_index);                                                             // movl _tls_index(%rip), %edx
+                      href.index:=reg_index; href.scalefactor:=8;
+                    end;
+                  current_asmdata.CurrAsmList.concat(taicpu.op_ref_reg(A_MOV,S_Q,href,hregister));         // movq (%rcx), %rcx  |  movq (%rcx,%rdx,8), %rcx
+
+                  reg_offset:=cg.getintregister(current_asmdata.CurrAsmList,OS_INT);
+                  href:=location.reference; // from inherited
+
+                if not (cs_asm_pre_binutils_2_25 in current_settings.globalswitches) then
+                  begin
+                    href.refaddr:=addr_secrel;
+                    current_asmdata.CurrAsmList.concat(taicpu.op_ref_reg(A_MOV,S_L,href,
+                      newreg(getregtype(reg_offset),getsupreg(reg_offset),R_SUBD)));                       // movl $threadvar@secrel32, %rax
+                  end
+                else // Old binutils (2.21) not support $threadvar@secrel32
+                  begin
+                    href.symbol.typ:=AT_DATA;
+                    href.refaddr:=addr_full;
+                    current_asmdata.CurrAsmList.concat(taicpu.op_ref_reg(A_MOV,S_Q,href,reg_offset));      // movq $threadvar, %rax
+                    reg_start:=cg.getintregister(current_asmdata.CurrAsmList,OS_INT);
+                    reference_reset_symbol(href,current_asmdata.RefAsmSymbol('___tls_start__',AT_DATA),0,sizeof(AInt),[]);
+                    href.refaddr:=addr_full;
+                    current_asmdata.CurrAsmList.concat(taicpu.op_ref_reg(A_MOV,S_Q,href,reg_start));       // movq $___tls_start__, %rdx
+                    current_asmdata.CurrAsmList.concat(taicpu.op_reg_reg(A_SUB,S_Q,reg_start,reg_offset)); // subq %rdx, %rax
+                  end;
+
+                  reference_reset(location.reference,location.reference.alignment,location.reference.volatility);
+                  location.reference.base:=hregister;
+                  location.reference.index:=reg_offset;                                                    // (%rcx,%rax)
+                end;
               else
                 ;
             end;
diff --git a/rtl/inc/heaptrc.pp b/rtl/inc/heaptrc.pp
index 920f448293..da7caade9f 100644
--- a/rtl/inc/heaptrc.pp
+++ b/rtl/inc/heaptrc.pp
@@ -1035,11 +1035,13 @@ var
    edata : ptruint; external name '__data_end__';
    sbss : ptruint; external name '__bss_start__';
    ebss : ptruint; external name '__bss_end__';
+{$ifndef FPC_SECTION_THREADVARS}
    TLSKey : PDWord; external name '_FPC_TlsKey';
    TLSSize : DWord; external name '_FPC_TlsSize';
 
 function TlsGetValue(dwTlsIndex : DWord) : pointer;
   {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'TlsGetValue';
+{$endif !FPC_SECTION_THREADVARS}
 {$endif}
 
 {$ifdef BEOS}
@@ -1104,6 +1106,7 @@ begin
   { inside data, rdata ... bss }
   if (ptruint(p)>=ptruint(@sdata)) and (ptruint(p)<ptruint(@ebss)) then
     exit;
+  {$ifndef FPC_SECTION_THREADVARS}
   { is program multi-threaded and p inside Threadvar range? }
   if TlsKey^<>-1 then
     begin
@@ -1112,6 +1115,7 @@ begin
           (ptruint(p)<ptruint(datap)+TlsSize)) then
         exit;
     end;
+  {$endif !FPC_SECTION_THREADVARS}
 {$endif windows}
 
 {$IFDEF OS2}
diff --git a/rtl/inc/thread.inc b/rtl/inc/thread.inc
index ce7ec796b4..07e20bb145 100644
--- a/rtl/inc/thread.inc
+++ b/rtl/inc/thread.inc
@@ -52,16 +52,20 @@ Var
           widestringmanager.ThreadInitProc;
 {$endif FPC_HAS_FEATURE_WIDESTRINGS}
 {$ifdef FPC_HAS_FEATURE_EXCEPTIONS}
+{$ifndef FPC_SECTION_THREADVARS}
         { ExceptAddrStack and ExceptObjectStack are threadvars       }
         { so every thread has its on exception handling capabilities }
         SysInitExceptions;
+{$endif !FPC_SECTION_THREADVARS}
 {$endif FPC_HAS_FEATURE_EXCEPTIONS}
 {$ifdef FPC_HAS_FEATURE_CONSOLEIO}
 {$ifndef EMBEDDED}
         { Open all stdio fds again }
         SysInitStdio;
+{$ifndef FPC_SECTION_THREADVARS}
         InOutRes:=0;
         // ErrNo:=0;
+{$endif !FPC_SECTION_THREADVARS}
 {$endif EMBEDDED}
 {$endif FPC_HAS_FEATURE_CONSOLEIO}
 {$ifdef FPC_HAS_FEATURE_STACKCHECK}
diff --git a/rtl/linux/si_impl.inc b/rtl/linux/si_impl.inc
index a727310551..e3ceee427e 100644
--- a/rtl/linux/si_impl.inc
+++ b/rtl/linux/si_impl.inc
@@ -20,7 +20,9 @@ procedure SysEntry_InitTLS(constref info: TEntryInformation); external name 'FPC
 
 var
   InitFinalTable : record end; external name 'INITFINAL';
+  {$ifndef FPC_SECTION_THREADVARS}
   ThreadvarTablesTable : record end; external name 'FPC_THREADVARTABLES';
+  {$endif}
   {$ifdef FPC_HAS_RESSTRINITS}
   ResStrInitTables : record end; external name 'FPC_RESSTRINITTABLES';
   {$endif}
diff --git a/rtl/win/sysosh.inc b/rtl/win/sysosh.inc
index 5bee0d4c17..4b65a1d41d 100644
--- a/rtl/win/sysosh.inc
+++ b/rtl/win/sysosh.inc
@@ -56,7 +56,9 @@ type
     {$ifdef WIN32}
     asm_exit : Procedure;stdcall;
     {$endif WIN32}
+    {$ifndef FPC_SECTION_THREADVARS}
     TlsKeyAddr : PDWord;
+    {$endif !FPC_SECTION_THREADVARS}
     SysInstance: {$ifdef CPU64}PQWord{$else}PLongInt{$endif};
     WideInitTables : Pointer;
   end;
@@ -117,7 +119,4 @@ var
 {$ifdef FPC_USE_WIN64_SEH}
 procedure _fpc_local_unwind(frame,target: Pointer);compilerproc;
 {$endif FPC_USE_WIN64_SEH}
-{$ifdef FPC_SECTION_THREADVARS}
-function fpc_tls_add(addr: pointer):pointer;compilerproc;
-{$endif FPC_SECTION_THREADVARS}
 
diff --git a/rtl/win/systhrd.inc b/rtl/win/systhrd.inc
index 10bb07fda4..dfa583fee7 100644
--- a/rtl/win/systhrd.inc
+++ b/rtl/win/systhrd.inc
@@ -19,6 +19,7 @@
                            Local WINApi imports
 *****************************************************************************}
 
+{$ifndef FPC_SECTION_THREADVARS}
 const
   { LocalAlloc flags  }
   LMEM_FIXED = 0;
@@ -34,15 +35,16 @@ function TlsGetValue(dwTlsIndex : DWord) : pointer;
   {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'TlsGetValue';
 function TlsSetValue(dwTlsIndex : DWord;lpTlsValue : pointer) : LongBool;
   {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'TlsSetValue';
+function LocalAlloc(uFlags:DWord; dwBytes:SIZE_T):Pointer;
+  {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'LocalAlloc';
+function LocalFree(hMem : Pointer):Pointer; {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'LocalFree';
+{$endif !FPC_SECTION_THREADVARS}
 function CreateThread(lpThreadAttributes : pointer;
   dwStackSize : SIZE_T; lpStartAddress : pointer;lpParameter : pointer;
   dwCreationFlags : DWord;var lpThreadId : DWord) : THandle;
   {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'CreateThread';
 procedure ExitThread(dwExitCode : DWord);
   {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'ExitThread';
-function LocalAlloc(uFlags:DWord; dwBytes:SIZE_T):Pointer;
-  {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'LocalAlloc';
-function LocalFree(hMem : Pointer):Pointer; {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'LocalFree';
 procedure Sleep(dwMilliseconds: DWord); {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'Sleep';
 function  WinSuspendThread (threadHandle : THandle) : dword; {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'SuspendThread';
 function  WinResumeThread  (threadHandle : THandle) : dword; {$ifdef wince}cdecl{$else}stdcall{$endif};external KernelDLL name 'ResumeThread';
@@ -98,6 +100,10 @@ var
                              Threadvar support
 *****************************************************************************}
 
+    var
+      MainThreadIdWin32 : DWORD;
+      
+{$ifndef FPC_SECTION_THREADVARS}
     var
       // public names are used by heaptrc unit
       threadvarblocksize : dword; public name '_FPC_TlsSize';
@@ -108,9 +114,6 @@ var
       TLSKey : PDWord = @TLSKeyVar; public name '_FPC_TlsKey';
       {$endif FPC_HAS_INDIRECT_ENTRY_INFORMATION}
 
-    var
-      MainThreadIdWin32 : DWORD;
-
     procedure SysInitThreadvar(var offset : dword;size : dword);
       begin
         offset:=threadvarblocksize;
@@ -199,6 +202,7 @@ var
             TlsSetValue(tlskey^, nil);
           end;
       end;
+{$endif !FPC_SECTION_THREADVARS}
 
 
 {*****************************************************************************
@@ -255,7 +259,9 @@ var
         writeln('Creating new thread');
 {$endif DEBUG_MT}
         { Initialize multithreading if not done }
+{$ifndef FPC_SECTION_THREADVARS}
         SysInitTLS;
+{$endif !FPC_SECTION_THREADVARS}
         IsMultiThread:=true;
 
         { the only way to pass data to the newly created thread
@@ -623,10 +629,17 @@ begin
     EnterCriticalSection   :=@SysEnterCriticalSection;
     TryEnterCriticalSection:=@SysTryEnterCriticalSection;
     LeaveCriticalSection   :=@SysLeaveCriticalSection;
+{$ifndef FPC_SECTION_THREADVARS}
     InitThreadVar          :=@SysInitThreadVar;
     RelocateThreadVar      :=@SysRelocateThreadVar;
     AllocateThreadVars     :=@SysAllocateThreadVars;
     ReleaseThreadVars      :=@SysReleaseThreadVars;
+{$else}
+    InitThreadVar          :=Nil;
+    RelocateThreadVar      :=Nil;
+    AllocateThreadVars     :=Nil;
+    ReleaseThreadVars      :=Nil;
+{$endif FPC_SECTION_THREADVARS}
     BasicEventCreate       :=@intBasicEventCreate;
     BasicEventDestroy      :=@intBasicEventDestroy;
     BasicEventResetEvent   :=@intBasicEventResetEvent;
@@ -641,10 +654,12 @@ begin
     end;
   SetThreadManager(WinThreadManager);
   ThreadID := GetCurrentThreadID;
+{$ifndef FPC_SECTION_THREADVARS}
 {$ifndef FPC_USE_TLS_DIRECTORY}
   if IsLibrary then
 {$endif}
     SysInitTLS;
+{$endif !FPC_SECTION_THREADVARS}
 
 {$ifndef WINCE}
   KernelHandle:=GetModuleHandle(KernelDLL);
diff --git a/rtl/win/systlsdir.inc b/rtl/win/systlsdir.inc
index db53b77a94..32e0f28e36 100644
--- a/rtl/win/systlsdir.inc
+++ b/rtl/win/systlsdir.inc
@@ -32,8 +32,10 @@ Const
   DLL_PROCESS_DETACH = 0;
   DLL_THREAD_DETACH = 3;
 
+{$ifndef FPC_SECTION_THREADVARS}
 var
    TlsKey : PDWord = @TlsKeyVar;
+{$endif !FPC_SECTION_THREADVARS}
 
 type
   TTlsDirectory=packed record
@@ -44,11 +46,13 @@ type
   end;
 
 
+{$ifndef FPC_SECTION_THREADVARS}
 function TlsGetValue(dwTlsIndex : DWord) : pointer; stdcall;
   external 'kernel32' name 'TlsGetValue';
 
 procedure InitSystemThreads; external name '_FPC_InitSystemThreads';
 procedure SysAllocateThreadVars; external name '_FPC_SysAllocateThreadVars';
+{$endif !FPC_SECTION_THREADVARS}
 procedure InitHeap; external name '_FPC_InitHeap';
 
 {$ifdef FPC_HAS_INDIRECT_ENTRY_INFORMATION}
@@ -58,10 +62,9 @@ procedure SetupEntryInformation(constref info: TEntryInformation); external name
 {$endif FPC_INSSIDE_SYSINIT}
 
 procedure Exec_Tls_callback(Handle : pointer; reason : Dword; Reserved : pointer);
-  stdcall; [public,alias:'_FPC_Tls_Callback'];
+  stdcall;
   begin
-     if IsLibrary then
-       Exit;
+     Assert(not IsLibrary);
      case reason of
        { For executables, DLL_PROCESS_ATTACH is called *before* the entry point,
          and DLL_PROCESS_DETACH is called *after* RTL shuts down and calls ExitProcess.
@@ -83,12 +86,19 @@ procedure Exec_Tls_callback(Handle : pointer; reason : Dword; Reserved : pointer
              the threadvar handling can be correctly initialized }
            SetupEntryInformation(SysInitEntryInformation);
            {$endif FPC_HAS_INDIRECT_ENTRY_INFORMATION}
+           { the DLL_PROCESS_ATTACH initialization for the heap is required, because
+             if the DLL_PROCESS_ATTACH entry of a used DLL creates a thread, DLL_THREAD_ATTACH
+             of the executable will be called so that the RTL for the thread can be initialized
+             and that needs the code in DLL_PROCESS_ATTACH to be run first }
            InitHeap;
+         {$ifndef FPC_SECTION_THREADVARS}
            InitSystemThreads;
+         {$endif !FPC_SECTION_THREADVARS}
          end;
 
        DLL_THREAD_ATTACH :
          begin
+         {$ifndef FPC_SECTION_THREADVARS}
          {  !!! SysInitMultithreading must NOT be called here. Windows guarantees that
             the main thread invokes PROCESS_ATTACH, not THREAD_ATTACH. So this always
             executes in non-main thread. SysInitMultithreading() here will cause
@@ -100,6 +110,7 @@ procedure Exec_Tls_callback(Handle : pointer; reason : Dword; Reserved : pointer
          }
            { Allocate Threadvars  }
            SysAllocateThreadVars;
+         {$endif !FPC_SECTION_THREADVARS}
 
            { NS : no idea what is correct to pass here - pass dummy value for now }
            { passing a dummy is ok, the correct value is read from the coff header of SysInstance (FK) }
@@ -107,7 +118,11 @@ procedure Exec_Tls_callback(Handle : pointer; reason : Dword; Reserved : pointer
          end;
        DLL_THREAD_DETACH :
          begin
+         {$ifdef FPC_SECTION_THREADVARS}
+           if ThreadID<>0 then
+         {$else !FPC_SECTION_THREADVARS}
            if TlsGetValue(TLSKey^)<>nil then
+         {$endif !FPC_SECTION_THREADVARS}
              DoneThread; { Assume everything is idempotent there }
          end;
      end;
diff --git a/rtl/win/syswin.inc b/rtl/win/syswin.inc
index 63ee752569..37fb574f33 100644
--- a/rtl/win/syswin.inc
+++ b/rtl/win/syswin.inc
@@ -408,10 +408,12 @@ function Dll_entry{$ifdef FPC_HAS_INDIRECT_ENTRY_INFORMATION}(constref info : TE
          end;
        DLL_THREAD_ATTACH :
          begin
+         {$ifndef FPC_SECTION_THREADVARS}
            { SysInitMultithreading must not be called here,
              see comments in exec_tls_callback below }
            { Allocate Threadvars  }
            SysAllocateThreadVars;
+         {$endif !FPC_SECTION_THREADVARS}
 
            { NS : no idea what is correct to pass here - pass dummy value for now }
            { passing a dummy is ok, the correct value is read from the coff header of SysInstance (FK) }
@@ -425,7 +427,11 @@ function Dll_entry{$ifdef FPC_HAS_INDIRECT_ENTRY_INFORMATION}(constref info : TE
            if assigned(Dll_Thread_Detach_Hook) then
              Dll_Thread_Detach_Hook(DllParam);
            { Release Threadvars }
+         {$ifdef FPC_SECTION_THREADVARS}
+           if ThreadID<>0 then
+         {$else !FPC_SECTION_THREADVARS}
            if TlsGetValue(TLSKey^)<>nil then
+         {$endif !FPC_SECTION_THREADVARS}
              DoneThread; { Assume everything is idempotent there }
          end;
        DLL_PROCESS_DETACH :
@@ -439,9 +445,11 @@ function Dll_entry{$ifdef FPC_HAS_INDIRECT_ENTRY_INFORMATION}(constref info : TE
                InternalExit;
              end;
 
+         {$ifndef FPC_SECTION_THREADVARS}
            SysReleaseThreadVars;
            { Free TLS resources used by ThreadVars }
            SysFiniMultiThreading;
+         {$endif !FPC_SECTION_THREADVARS}
            MainThreadIDWin32:=0;
          end;
      end;
diff --git a/rtl/win32/sysinit.inc b/rtl/win32/sysinit.inc
index 0eea501af0..c8f7b6f042 100644
--- a/rtl/win32/sysinit.inc
+++ b/rtl/win32/sysinit.inc
@@ -16,10 +16,14 @@
 
    var
       SysInstance : LongInt;
+ {$ifndef FPC_SECTION_THREADVARS}
       TlsKeyVar: DWord = $ffffffff;
+ {$endif !FPC_SECTION_THREADVARS}
 
       InitFinalTable : record end; external name 'INITFINAL';
+ {$ifndef FPC_SECTION_THREADVARS}
       ThreadvarTablesTable : record end; external name 'FPC_THREADVARTABLES';
+ {$endif !FPC_SECTION_THREADVARS}
       WideInitTables : record end; external name 'FPC_WIDEINITTABLES';
  {$ifdef FPC_HAS_RESSTRINITS}
       ResStrInitTables : record end; external name 'FPC_RESSTRINITTABLES';
@@ -53,7 +57,11 @@
       STD_INPUT_HANDLE = dword(-10);
       SysInitEntryInformation : TEntryInformation = (
         InitFinalTable : @InitFinalTable;
+{$ifdef FPC_SECTION_THREADVARS}
+        ThreadvarTablesTable : nil;
+{$else FPC_SECTION_THREADVARS}
         ThreadvarTablesTable : @ThreadvarTablesTable;
+{$endif FPC_SECTION_THREADVARS}
         ResourceStringTables : @ResourceStringTables;
 {$ifdef FPC_HAS_RESSTRINITS}
         ResStrInitTables : @ResStrInitTables;
@@ -65,7 +73,9 @@
         valgrind_used : false;
         OS : (
           asm_exit : @asm_exit;
+{$ifndef FPC_SECTION_THREADVARS}
           TlsKeyAddr : @TlsKeyVar;
+{$endif !FPC_SECTION_THREADVARS}
           SysInstance : @SysInstance;
           WideInitTables: @WideInitTables;
           );
diff --git a/rtl/win32/system.pp b/rtl/win32/system.pp
index 4ffa0b65ff..4cf69de94d 100644
--- a/rtl/win32/system.pp
+++ b/rtl/win32/system.pp
@@ -78,7 +78,9 @@ end;
 
 procedure OsSetupEntryInformation(constref info: TEntryInformation);
 begin
+{$ifndef FPC_SECTION_THREADVARS}
   TlsKey := info.OS.TlsKeyAddr;
+{$endif !FPC_SECTION_THREADVARS}
   FPCSysInstance := info.OS.SysInstance;
   WStrInitTablesTable := info.OS.WideInitTables;
 end;
@@ -498,41 +500,6 @@ begin
 end;
 {$endif Set_i386_Exception_handler}
 
-{$ifdef FPC_SECTION_THREADVARS}
-function fpc_tls_add(addr: pointer): pointer; assembler; nostackframe;
-  [public,alias: 'FPC_TLS_ADD']; compilerproc;
-  asm
-      sub   $tls_data_start,%eax
-      cmpb  $0,IsLibrary
-      mov   _tls_index,%ecx
-      jnz   .L1
-      mov   %fs:(0x2c),%edx
-      add   (%edx,%ecx,4),%eax
-      ret
-.L1:
-      push  %ebx
-      mov   %eax,%ebx
-      call  GetLastError
-      push  %eax                      { save LastError }
-      push  _tls_index
-      call  TlsGetValue
-      test  %eax,%eax
-      jnz   .L2
-      { This can happen when a thread existed before DLL was loaded,
-        or if DisableThreadLibraryCalls was called. }
-      call  SysAllocateThreadVars
-      mov   $0x1000000,%eax
-      call  InitThread
-      push  _tls_index
-      call  TlsGetValue
-.L2:
-      add   %eax,%ebx
-      call  SetLastError              { restore (value is on stack) }
-      mov   %ebx,%eax
-      pop   %ebx
-  end;
-{$endif FPC_SECTION_THREADVARS}
-
 function CheckInitialStkLen(stklen : SizeUInt) : SizeUInt;
   type
     tdosheader = packed record
diff --git a/rtl/win64/sysinit.pp b/rtl/win64/sysinit.pp
index 5d396beb8f..f658cccb2b 100644
--- a/rtl/win64/sysinit.pp
+++ b/rtl/win64/sysinit.pp
@@ -21,10 +21,14 @@ unit sysinit;
 
    var
       SysInstance : QWord;
+{$ifndef FPC_SECTION_THREADVARS}
       TlsKeyVar: DWord = $ffffffff;
+{$endif !FPC_SECTION_THREADVARS}
 
       InitFinalTable : record end; external name 'INITFINAL';
+{$ifndef FPC_SECTION_THREADVARS}
       ThreadvarTablesTable : record end; external name 'FPC_THREADVARTABLES';
+{$endif !FPC_SECTION_THREADVARS}
       WideInitTables : record end; external name 'FPC_WIDEINITTABLES';
 {$ifdef FPC_HAS_RESSTRINITS}
       ResStrInitTables : record end; external name 'FPC_RESSTRINITTABLES';
@@ -58,7 +62,11 @@ unit sysinit;
       STD_INPUT_HANDLE = dword(-10);
       SysInitEntryInformation : TEntryInformation = (
         InitFinalTable : @InitFinalTable;
+{$ifdef FPC_SECTION_THREADVARS}
+        ThreadvarTablesTable : nil;
+{$else FPC_SECTION_THREADVARS}
         ThreadvarTablesTable : @ThreadvarTablesTable;
+{$endif FPC_SECTION_THREADVARS}
         ResourceStringTables : @ResourceStringTables;
 {$ifdef FPC_HAS_RESSTRINITS}
         ResStrInitTables : @ResStrInitTables;
@@ -69,7 +77,9 @@ unit sysinit;
         PascalMain : @PascalMain;
         valgrind_used : false;
         OS : (
+{$ifndef FPC_SECTION_THREADVARS}
           TlsKeyAddr : @TlsKeyVar;
+{$endif !FPC_SECTION_THREADVARS}
           SysInstance : @SysInstance;
           WideInitTables: @WideInitTables;
           );
@@ -124,6 +134,9 @@ unit sysinit;
       sysinstance:=_hinstance;
       dllreason:=_dllreason;
       dllparam:=PtrInt(_dllparam);
+{$if defined(FPC_USE_TLS_DIRECTORY) and defined(FPC_SECTION_THREADVARS)}
+      LinkIn(@tlsdir,@tls_callback_end,nil);
+{$endif}
       SetupEntryInformation;
       DLL_Entry(SysInitEntryInformation);
     end;
@@ -135,6 +148,9 @@ unit sysinit;
       sysinstance:=_hinstance;
       dllreason:=_dllreason;
       dllparam:=PtrInt(_dllparam);
+{$if defined(FPC_USE_TLS_DIRECTORY) and defined(FPC_SECTION_THREADVARS)}
+      LinkIn(@tlsdir,@tls_callback_end,nil);
+{$endif}
       SetupEntryInformation;
       DLL_Entry(SysInitEntryInformation);
     end;
diff --git a/rtl/win64/system.pp b/rtl/win64/system.pp
index 6cf7b174cd..a5771e281e 100644
--- a/rtl/win64/system.pp
+++ b/rtl/win64/system.pp
@@ -122,7 +122,9 @@ procedure PascalMain;external name 'PASCALMAIN';
 
 procedure OsSetupEntryInformation(constref info: TEntryInformation);
 begin
+{$ifndef FPC_SECTION_THREADVARS}
   TlsKey := info.OS.TlsKeyAddr;
+{$endif !FPC_SECTION_THREADVARS}
   FPCSysInstance := info.OS.SysInstance;
   WStrInitTablesTable := info.OS.WideInitTables;
 end;
@@ -504,47 +506,6 @@ begin
 end;
 {$endif VER3_0}
 
-{$ifdef FPC_SECTION_THREADVARS}
-function fpc_tls_add(addr: pointer): pointer; assembler; nostackframe;
-  [public,alias: 'FPC_TLS_ADD']; compilerproc;
-  asm
-      sub   $56,%rsp                  { 32 spill area + 16 local vars + 8 misalignment }
-  .seh_stackalloc 56
-  .seh_endprologue
-      lea   tls_data_start(%rip),%rax
-      sub   %rax,%rcx
-      cmpb  $0,IsLibrary(%rip)
-      mov   _tls_index(%rip),%eax
-      jnz   .L1
-      mov   %gs:(88),%rdx
-      add   (%rdx,%rax,8),%rcx
-      mov   %rcx,%rax
-      jmp   .L3
-.L1:
-      mov   %rcx,32(%rsp)
-      call  GetLastError
-      mov   %rax,40(%rsp)             { save LastError }
-      mov   _tls_index(%rip),%ecx
-      call  TlsGetValue
-      test  %rax,%rax
-      jnz   .L2
-      { This can happen when a thread existed before DLL was loaded,
-        or if DisableThreadLibraryCalls was called. }
-      call  SysAllocateThreadVars
-      mov   $0x1000000,%rcx
-      call  InitThread
-      mov   _tls_index(%rip),%ecx
-      call  TlsGetValue
-.L2:
-      add   %rax,32(%rsp)
-      mov   40(%rsp),%rcx
-      call  SetLastError
-      mov   32(%rsp),%rax
-.L3:
-      add   $56,%rsp
-  end;
-{$endif FPC_SECTION_THREADVARS}
-
 function CheckInitialStkLen(stklen : SizeUInt) : SizeUInt;
   type
     tdosheader = packed record
@@ -632,15 +593,19 @@ initialization
     InitHeap;
     InitSystemThreads;
   end;
+{$ifndef FPC_SECTION_THREADVARS}
   SysInitExceptions;
+{$endif !FPC_SECTION_THREADVARS}
   initunicodestringmanager;
   InitWin32Widestrings;
   SysInitStdIO;
   { Arguments }
   setup_arguments;
   InitSystemDynLibs;
+{$ifndef FPC_SECTION_THREADVARS}
   { Reset IO Error }
   InOutRes:=0;
+{$endif !FPC_SECTION_THREADVARS}
   ProcessID := GetCurrentProcessID;
   DispCallByIDProc:=@DoDispCallByIDError;
 
-- 
2.26.2.windows.1