diff --git a/CMakeLists.txt b/CMakeLists.txt
index 12dc810b..00fd4a01 100755
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -37,6 +37,7 @@ SET(ELFLOADER_SRC
     src/elfload_dump.c
     src/librarian.c
     src/stack.c
+    src/threads.c
     src/wrapper.c
     src/x86emu.c
     src/x86run.c
@@ -50,7 +51,7 @@ SET(ELFLOADER_SRC
 )
 
 add_executable(box86 ${ELFLOADER_SRC})
-target_link_libraries(box86 m dl rt)
+target_link_libraries(box86 m dl rt pthread)
 
 add_test(test01 ${CMAKE_COMMAND} -D TEST_PROGRAM=${CMAKE_BINARY_DIR}/box86 
     -D TEST_ARGS=${CMAKE_SOURCE_DIR}/tests/test01 -D TEST_OUTPUT=tmpfile.txt 
@@ -75,4 +76,9 @@ add_test(test04 ${CMAKE_COMMAND} -D TEST_PROGRAM=${CMAKE_BINARY_DIR}/box86
 add_test(test05 ${CMAKE_COMMAND} -D TEST_PROGRAM=${CMAKE_BINARY_DIR}/box86 
     -D TEST_ARGS=${CMAKE_SOURCE_DIR}/tests/test05 -D TEST_ARGS2=7 -D TEST_OUTPUT=tmpfile.txt 
     -D TEST_REFERENCE=${CMAKE_SOURCE_DIR}/tests/ref05.txt
+    -P ${CMAKE_SOURCE_DIR}/runTest.cmake )
+
+add_test(test06 ${CMAKE_COMMAND} -D TEST_PROGRAM=${CMAKE_BINARY_DIR}/box86 
+    -D TEST_ARGS=${CMAKE_SOURCE_DIR}/tests/test06 -D TEST_OUTPUT=tmpfile.txt 
+    -D TEST_REFERENCE=${CMAKE_SOURCE_DIR}/tests/ref06.txt
     -P ${CMAKE_SOURCE_DIR}/runTest.cmake )
\ No newline at end of file
diff --git a/src/box86context.h b/src/box86context.h
index 2c51a286..632eb869 100755
--- a/src/box86context.h
+++ b/src/box86context.h
@@ -8,6 +8,7 @@ typedef struct elfheader_s elfheader_t;
 typedef struct x86emu_s x86emu_t;
 typedef struct zydis_s zydis_t;
 typedef struct lib_s lib_t;
+typedef struct bridge_s bridge_t;
 
 typedef struct box86context_s {
     path_collection_t   box86_path;     // PATH env. variable
@@ -36,6 +37,8 @@ typedef struct box86context_s {
 
     lib_t               *maplib;        // lib and symbols handling
 
+    bridge_t            *threads;       // threads
+
 } box86context_t;
 
 box86context_t *NewBox86Context(int argc);
diff --git a/src/bridge.c b/src/bridge.c
index 0f8b5d67..95b28282 100755
--- a/src/bridge.c
+++ b/src/bridge.c
@@ -3,18 +3,9 @@
 #include <stdint.h>
 
 #include "bridge.h"
+#include "bridge_private.h"
 #include "wrapper.h"
 
-#pragma pack(push, 1)
-typedef struct onebridge_s {
-    uint8_t CC;     // CC int 0x3
-    uint8_t S, C;   // 'S' 'C', just a signature
-    wrapper_t w;    // wrapper
-    uintptr_t f;    // the function for the wrapper
-    uint8_t C3;     // C3 ret
-} onebridge_t;
-#pragma pack(pop)
-
 #define NBRICK  16
 typedef struct brick_s brick_t;
 typedef struct brick_s {
diff --git a/src/bridge_private.h b/src/bridge_private.h
new file mode 100755
index 00000000..4db190a5
--- /dev/null
+++ b/src/bridge_private.h
@@ -0,0 +1,17 @@
+#ifndef __BRIDGE_PRIVATE_H_
+#define __BRIDGE_PRIVATE_H_
+#include <stdint.h>
+
+#include "wrapper.h"
+
+#pragma pack(push, 1)
+typedef struct onebridge_s {
+    uint8_t CC;     // CC int 0x3
+    uint8_t S, C;   // 'S' 'C', just a signature
+    wrapper_t w;    // wrapper
+    uintptr_t f;    // the function for the wrapper
+    uint8_t C3;     // C3 ret
+} onebridge_t;
+#pragma pack(pop)
+
+#endif //__BRIDGE_PRIVATE_H_
\ No newline at end of file
diff --git a/src/elfloader.c b/src/elfloader.c
index 8eefb77f..b5ec6b73 100755
--- a/src/elfloader.c
+++ b/src/elfloader.c
@@ -301,8 +301,9 @@ void AddGlobalsSymbols(lib_t* maplib, elfheader_t* h)
         if(((h->SymTab[i].st_info == 18) || h->SymTab[i].st_info == 17) && (h->SymTab[i].st_other==0) && (h->SymTab[i].st_shndx!=0)) {
             const char * symname = h->StrTab+h->SymTab[i].st_name;
             uintptr_t offs = h->SymTab[i].st_value + h->delta;
-            printf_log(LOG_DUMP, "Adding Symbol \"%s\" with offset=%p\n", symname, offs);
-            AddSymbol(maplib, symname, offs);
+            uint32_t sz = h->SymTab[i].st_size;
+            printf_log(LOG_DUMP, "Adding Symbol \"%s\" with offset=%p sz=%d\n", symname, offs, sz);
+            AddSymbol(maplib, symname, offs, sz);
         }
     }
 }
diff --git a/src/librarian.c b/src/librarian.c
index a4bfd672..0b9d73aa 100755
--- a/src/librarian.c
+++ b/src/librarian.c
@@ -1,6 +1,8 @@
 #include <stdio.h>
 #include <stdlib.h>
 
+#include <pthread.h>
+
 #include "debug.h"
 #include "librarian.h"
 #include "librarian_private.h"
@@ -39,11 +41,13 @@ int32_t my__libc_start_main(x86emu_t* emu, int *(main) (int, char * *, char * *)
     int argc, char * * ubp_av, void (*init) (void), void (*fini) (void), 
     void (*rtld_fini) (void), void (* stack_end)); // implemented in x86run_private.c
 uint32_t LibSyscall(x86emu_t *emu); // implemented in x86syscall.c
+int my_pthread_create(x86emu_t *emu, void* t, void* attr, void* start_routine, void* arg); //implemented in thread.c
 
 uintptr_t CreateSymbol(lib_t *maplib, const char* name)
 {
     // look for symbols that can be created
     uintptr_t addr = 0;
+    // libc
     if(strcmp(name, "__stack_chk_fail")==0) {
         addr = AddBridge(maplib->bridge, vFE, my__stack_chk_fail);
     } else if(strcmp(name, "__libc_start_main")==0) {
@@ -64,17 +68,30 @@ uintptr_t CreateSymbol(lib_t *maplib, const char* name)
         addr = AddBridge(maplib->bridge, iFi, putchar);
     } else if(strcmp(name, "strtol")==0) {
         addr = AddBridge(maplib->bridge, iFppi, strtol);
+    } else if(strcmp(name, "strerror")==0) {
+        addr = AddBridge(maplib->bridge, pFv, strerror);
+    } //pthread
+    else if(strcmp(name, "pthread_self")==0) {
+        addr = AddBridge(maplib->bridge, uFv, pthread_self);
+    } else if(strcmp(name, "pthread_create")==0) {
+        addr = AddBridge(maplib->bridge, iFEpppp, my_pthread_create);
+    } else if(strcmp(name, "pthread_equal")==0) {
+        addr = AddBridge(maplib->bridge, iFuu, pthread_equal);
+    } else if(strcmp(name, "pthread_join")==0) {
+        addr = AddBridge(maplib->bridge, iFup, pthread_join);
     }
+    
     if(addr)
-        AddSymbol(maplib, name, addr);
+        AddSymbol(maplib, name, addr, 12);
     return addr;
 }
 
-void AddSymbol(lib_t *maplib, const char* name, uintptr_t addr)
+void AddSymbol(lib_t *maplib, const char* name, uintptr_t addr, uint32_t sz)
 {
     int ret;
     khint_t k = kh_put(maplib_t, maplib->maplib, name, &ret);
     kh_value(maplib->maplib, k).offs = addr;
+    kh_value(maplib->maplib, k).sz = sz;
 }
 uintptr_t FindSymbol(lib_t *maplib, const char* name)
 {
@@ -83,3 +100,13 @@ uintptr_t FindSymbol(lib_t *maplib, const char* name)
         return CreateSymbol(maplib, name);
     return kh_val(maplib->maplib, k).offs;
 }
+
+int GetSymbolStartEnd(lib_t* maplib, const char* name, uintptr_t* start, uintptr_t* end)
+{
+    khint_t k = kh_get(maplib_t, maplib->maplib, name);
+    if(k==kh_end(maplib->maplib))
+        return 0;
+    *start = kh_val(maplib->maplib, k).offs;
+    *end = *start + kh_val(maplib->maplib, k).sz;
+    return 1;
+}
\ No newline at end of file
diff --git a/src/librarian.h b/src/librarian.h
index 1b5ae421..61394bb4 100755
--- a/src/librarian.h
+++ b/src/librarian.h
@@ -8,7 +8,8 @@ typedef struct bridge_s bridge_t;
 lib_t *NewLibrarian();
 void FreeLibrarian(lib_t **maplib);
 
-void AddSymbol(lib_t *maplib, const char* name, uintptr_t addr);
+void AddSymbol(lib_t *maplib, const char* name, uintptr_t addr, uint32_t sz);
 uintptr_t FindSymbol(lib_t *maplib, const char* name);
+int GetSymbolStartEnd(lib_t* maplib, const char* name, uintptr_t* start, uintptr_t* end);
 
 #endif //__LIBRARIAN_H_
\ No newline at end of file
diff --git a/src/librarian_private.h b/src/librarian_private.h
index 9a046ac7..b5543408 100755
--- a/src/librarian_private.h
+++ b/src/librarian_private.h
@@ -5,6 +5,7 @@
 
 typedef struct {
     uintptr_t   offs;
+    uint32_t    sz;
     // need to track type of symbol?
     // need to track origin?
 } onelib_t;
diff --git a/src/main.c b/src/main.c
index 82f77dc3..3484bf82 100755
--- a/src/main.c
+++ b/src/main.c
@@ -129,7 +129,7 @@ int main(int argc, const char **argv, const char **env) {
     // init random seed
     srandom(time(NULL));
 
-    // check BOX86_loG debug level
+    // check BOX86_LOG debug level
     LoadLogEnv();
     
     // Create a new context
@@ -153,8 +153,7 @@ int main(int argc, const char **argv, const char **env) {
 
     p = getenv("BOX86_TRACE");
     if(p) {
-        setbuf(stdout, NULL);
-        if (strcmp(p, "1")==0)
+        if (strcmp(p, "0"))
             context->x86trace = 1;
     }
     if(context->x86trace) {
@@ -243,9 +242,20 @@ int main(int argc, const char **argv, const char **env) {
     // setup the stack...
     Push(context->emu, (uint32_t)context->argv);
     Push(context->emu, context->argc);
-    SetupX86Emu(context->emu);
+    SetupX86Emu(context->emu, NULL, NULL);
     SetEAX(context->emu, context->argc);
     SetEBX(context->emu, (uint32_t)context->argv);
+
+    p = getenv("BOX86_TRACE");
+    if(p) {
+        setbuf(stdout, NULL);
+        uintptr_t trace_start, trace_end;
+        if (strcmp(p, "1")==0)
+            SetTraceEmu(context->emu, 0, 0);
+        else if (GetSymbolStartEnd(context->maplib, p, &trace_start, &trace_end))
+            SetTraceEmu(context->emu, trace_start, trace_end);
+    }
+
     // emulate!
     printf_log(LOG_DEBUG, "Start x86emu on Main\n");
     Run(context->emu);
diff --git a/src/threads.c b/src/threads.c
new file mode 100755
index 00000000..c918a301
--- /dev/null
+++ b/src/threads.c
@@ -0,0 +1,57 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <pthread.h>
+
+#include "box86context.h"
+#include "threads.h"
+#include "x86emu_private.h"
+#include "bridge_private.h"
+#include "x86run.h"
+#include "x86emu.h"
+#include "stack.h"
+
+// memory handling to be perfected...
+// keep a hash thread_t -> emu to set emu->quit to 1 on pthread_cancel
+
+typedef struct emuthread_s {
+	x86emu_t 	*emu;
+	onebridge_t	routine;
+	int			stacksize;
+	void		*stack;
+} emuthread_t;
+
+void* pthread_routine(void* p)
+{
+	emuthread_t *et = (emuthread_t*)p;
+	Run(et->emu);
+	void* r = (void*)GetEAX(et->emu);
+	FreeX86Emu(&et->emu);
+	free(et->stack);
+	free(et);
+	return r;
+}
+
+int my_pthread_create(x86emu_t *emu, void* t, void* attr, void* start_routine, void* arg)
+{
+	emuthread_t *emuthread = (emuthread_t*)calloc(1, sizeof(emuthread_t));
+	emuthread->routine.CC = 0xCC;
+	emuthread->routine.S = 'S'; emuthread->routine.C = 'C';
+	emuthread->routine.w = pFp;
+	emuthread->routine.f = (uintptr_t)start_routine;
+	emuthread->routine.C3 = 0xC3;
+
+	emuthread->stacksize = 2*1024*1024;	//default stack size is 2Mo
+	// TODO: get stack size inside attr
+	emuthread->stack = calloc(1, emuthread->stacksize);
+	emuthread->emu = NewX86Emu(emu->context, (uintptr_t)start_routine, (uintptr_t)emuthread->stack, 
+		emuthread->stacksize);
+	SetupX86Emu(emuthread->emu, emu->shared_global, emu->globals);
+	emuthread->emu->trace_start = emu->trace_start;
+	emuthread->emu->trace_end = emu->trace_end;
+	Push(emuthread->emu, (uintptr_t)arg);
+	PushExit(emuthread->emu);
+
+	// create thread
+	return pthread_create((pthread_t*)t, (const pthread_attr_t *)attr, 
+		pthread_routine, emuthread);
+}
\ No newline at end of file
diff --git a/src/threads.h b/src/threads.h
new file mode 100755
index 00000000..08357ea3
--- /dev/null
+++ b/src/threads.h
@@ -0,0 +1,7 @@
+#ifndef _THREADS_H_
+#define _THREADS_H_
+
+#include "bridge.h"
+
+
+#endif //_THREADS_H_
\ No newline at end of file
diff --git a/src/wrapper.c b/src/wrapper.c
index 0af9d80c..53a2a716 100755
--- a/src/wrapper.c
+++ b/src/wrapper.c
@@ -23,15 +23,23 @@ typedef void        (*vFv_t)();
 typedef void        (*vFp_t)(void*);
 typedef void        (*vFE_t)(x86emu_t*);
 typedef int32_t     (*iFv_t)();
+typedef uint32_t    (*uFv_t)();
+typedef void*       (*pFv_t)();
 typedef uint32_t    (*uFE_t)(x86emu_t*);
 typedef void        (*vFi_t)(int32_t);
 typedef int32_t     (*iFi_t)(int32_t);
 typedef int32_t     (*iFp_t)(void*);
+typedef void*       (*pFp_t)(void*);
 typedef int32_t     (*iFpp_t)(void*, void*);
+typedef int32_t     (*iFii_t)(int32_t, int32_t);
+typedef int32_t     (*iFip_t)(int32_t, void*);
+typedef int32_t     (*iFuu_t)(uint32_t, uint32_t);
+typedef int32_t     (*iFup_t)(uint32_t, void*);
 typedef int32_t     (*iFppi_t)(void*, void*, int32_t);
 typedef void*       (*pFuu_t)(uint32_t, uint32_t);
 typedef int32_t     (*iFipp_t)(int, void*, void*);
 typedef int32_t     (*iFppp_t)(void*, void*, void*);
+typedef int32_t     (*iFEpppp_t)(x86emu_t*, void*, void*, void*, void*);
 typedef int32_t     (*iFEpippppp_t)(x86emu_t*, void*, int32_t, void*, void*, void*, void*, void*);
 
 #define DEF(A) A f = (A)fnc
@@ -54,7 +62,12 @@ void    vFE(x86emu_t *emu, uintptr_t fnc)
 void    iFv(x86emu_t *emu, uintptr_t fnc)
 {
     DEF(iFv_t);
-    *(int32_t*)&R_EAX = f();
+    R_EAX = *(uint32_t*)f();
+}
+void    uFv(x86emu_t *emu, uintptr_t fnc)
+{
+    DEF(uFv_t);
+    R_EAX = f();
 }
 void    uFE(x86emu_t *emu, uintptr_t fnc)
 {
@@ -66,6 +79,11 @@ void    vFi(x86emu_t *emu, uintptr_t fnc)
     DEF(vFi_t);
     f(i32(0));
 }
+void    pFv(x86emu_t *emu, uintptr_t fnc)
+{
+    DEF(pFv_t);
+    R_EAX = *(uint32_t*)f();
+}
 void    iFi(x86emu_t *emu, uintptr_t fnc)
 {
     DEF(iFi_t);
@@ -76,6 +94,11 @@ void    iFp(x86emu_t *emu, uintptr_t fnc)
     DEF(iFp_t);
     R_EAX = (uint32_t)f(p(0));
 }
+void    pFp(x86emu_t *emu, uintptr_t fnc)
+{
+    DEF(pFp_t);
+    R_EAX = (uint32_t)f(p(0));
+}
 void    iFpp(x86emu_t *emu, uintptr_t fnc)
 {
     DEF(iFpp_t);
@@ -91,6 +114,26 @@ void    pFuu(x86emu_t *emu, uintptr_t fnc)
     DEF(pFuu_t);
     R_EAX = (uintptr_t)f(u32(0), u32(4));
 }
+void    iFii(x86emu_t *emu, uintptr_t fnc)
+{
+    DEF(iFii_t);
+    R_EAX = (uint32_t)f(i32(0), i32(4));
+}
+void    iFip(x86emu_t *emu, uintptr_t fnc)
+{
+    DEF(iFip_t);
+    R_EAX = (uint32_t)f(i32(0), p(4));
+}
+void    iFuu(x86emu_t *emu, uintptr_t fnc)
+{
+    DEF(iFuu_t);
+    R_EAX = (uint32_t)f(u32(0), u32(4));
+}
+void    iFup(x86emu_t *emu, uintptr_t fnc)
+{
+    DEF(iFup_t);
+    R_EAX = (uint32_t)f(u32(0), p(4));
+}
 void    iFpV(x86emu_t *emu, uintptr_t fnc)
 {
     DEF(iFpp_t);
@@ -111,6 +154,11 @@ void    iFvopV(x86emu_t *emu, uintptr_t fnc)
     DEF(iFppp_t);
     R_EAX = (uint32_t)f((void*)stdout, p(4), (void*)stack(8));
 }
+void	iFEpppp(x86emu_t *emu, uintptr_t fnc)
+{
+    DEF(iFEpppp_t);
+    R_EAX = (uint32_t)f(emu, p(0), p(4), p(8), p(12));
+}
 void    iFEpippppp(x86emu_t *emu, uintptr_t fnc)
 {
     DEF(iFEpippppp_t);
diff --git a/src/wrapper.h b/src/wrapper.h
index 89e1860c..d30f555a 100755
--- a/src/wrapper.h
+++ b/src/wrapper.h
@@ -20,16 +20,24 @@ void    vFp(x86emu_t *emu, uintptr_t fnc);
 void    vFE(x86emu_t *emu, uintptr_t fnc);
 void    uFE(x86emu_t *emu, uintptr_t fnc);
 void    iFv(x86emu_t *emu, uintptr_t fnc);
+void    uFv(x86emu_t *emu, uintptr_t fnc);
+void    pFv(x86emu_t *emu, uintptr_t fnc);
 void    vFi(x86emu_t *emu, uintptr_t fnc);
 void    iFi(x86emu_t *emu, uintptr_t fnc);
 void    iFp(x86emu_t *emu, uintptr_t fnc);
+void    pFp(x86emu_t *emu, uintptr_t fnc);
 void    iFpp(x86emu_t *emu, uintptr_t fnc);
 void    iFppi(x86emu_t *emu, uintptr_t fnc);
 void    pFuu(x86emu_t *emu, uintptr_t fnc);
+void    iFii(x86emu_t *emu, uintptr_t fnc);
+void    iFip(x86emu_t *emu, uintptr_t fnc);
+void    iFuu(x86emu_t *emu, uintptr_t fnc);
+void    iFup(x86emu_t *emu, uintptr_t fnc);
 void    iFpv(x86emu_t *emu, uintptr_t fnc);
 void    iF1pV(x86emu_t *emu, uintptr_t fnc);
 void    iFopV(x86emu_t *emu, uintptr_t fnc);
 void    iFvopV(x86emu_t *emu, uintptr_t fnc);
+void	iFEpppp(x86emu_t *emu, uintptr_t fnc);
 void    iFEpippppp(x86emu_t *emu, uintptr_t fnc);   // this is __libc_start_main basically
 
 
diff --git a/src/x86emu.c b/src/x86emu.c
index 4d6a0463..ee016af7 100755
--- a/src/x86emu.c
+++ b/src/x86emu.c
@@ -51,20 +51,34 @@ x86emu_t *NewX86Emu(box86context_t *context, uintptr_t start, uintptr_t stack, i
     return emu;
 }
 
-void SetupX86Emu(x86emu_t *emu)
+void SetupX86Emu(x86emu_t *emu, int* shared_global, void* globals)
 {
     printf_log(LOG_DEBUG, "Setup X86 Emu\n");
 
     // push "end emu" marker address
     PushExit(emu);
     // Setup the GS segment:
-    emu->globals = calloc(1, 256);  // arbitrary 256 byte size?
-    // calc canary...
-    uint8_t canary[4];
-    for (int i=0; i<4; ++i) canary[i] = 1 +  getrand(255);
-    canary[getrand(4)] = 0;
-    memcpy(emu->globals+0x14, canary, sizeof(canary));  // put canary in place
-    printf_log(LOG_DEBUG, "Setting up canary (for Stack protector) at GS:0x14, value:%08X\n", *(uint32_t*)canary);
+    if(shared_global) {
+        emu->globals = globals;
+        emu->shared_global = shared_global;
+    } else {
+        emu->globals = calloc(1, 256);  // arbitrary 256 byte size?
+        // calc canary...
+        uint8_t canary[4];
+        for (int i=0; i<4; ++i) canary[i] = 1 +  getrand(255);
+        canary[getrand(4)] = 0;
+        memcpy(emu->globals+0x14, canary, sizeof(canary));  // put canary in place
+        printf_log(LOG_DEBUG, "Setting up canary (for Stack protector) at GS:0x14, value:%08X\n", *(uint32_t*)canary);
+        emu->shared_global = (int*)calloc(1, sizeof(int));
+    }
+    (*emu->shared_global)++;
+}
+
+void SetTraceEmu(x86emu_t *emu, uintptr_t trace_start, uintptr_t trace_end)
+{
+    printf_log(LOG_INFO, "Setting trace only between %p and %p\n", trace_start, trace_end);
+    emu->trace_start = trace_start;
+    emu->trace_end = trace_end;
 }
 
 void FreeX86Emu(x86emu_t **x86emu)
@@ -74,8 +88,11 @@ void FreeX86Emu(x86emu_t **x86emu)
     printf_log(LOG_DEBUG, "Free a X86 Emu (%p)\n", *x86emu);
     if((*x86emu)->dec)
         DeleteX86TraceDecoder(&(*x86emu)->dec);
-    if((*x86emu)->globals)
-        free((*x86emu)->globals);
+    if((*x86emu)->shared_global && !(*(*x86emu)->shared_global)--) {
+        if((*x86emu)->globals)
+            free((*x86emu)->globals);
+        free((*x86emu)->shared_global);
+    }
     free(*x86emu);
     *x86emu = NULL;
 }
diff --git a/src/x86emu.h b/src/x86emu.h
index 7641b6d8..c23e003e 100755
--- a/src/x86emu.h
+++ b/src/x86emu.h
@@ -5,7 +5,8 @@ typedef struct x86emu_s x86emu_t;
 typedef struct box86context_s box86context_t;
 
 x86emu_t *NewX86Emu(box86context_t *context, uintptr_t start, uintptr_t stack, int stacksize);
-void SetupX86Emu(x86emu_t *emu);
+void SetupX86Emu(x86emu_t *emu, int* shared_gloabl, void* globals);
+void SetTraceEmu(x86emu_t *emu, uintptr_t trace_start, uintptr_t trace_end);
 void FreeX86Emu(x86emu_t **x86emu);
 
 uint32_t GetEAX(x86emu_t *emu);
diff --git a/src/x86emu_private.h b/src/x86emu_private.h
index e335aa22..7a239d25 100755
--- a/src/x86emu_private.h
+++ b/src/x86emu_private.h
@@ -34,8 +34,10 @@ typedef struct x86emu_s {
     int         error;
     // trace
     zydis_dec_t *dec;
+    uintptr_t   trace_start, trace_end;
     // global stuffs, pointed with GS: segment
     void        *globals;
+    int         *shared_global;
     // parent context
     box86context_t *context;
 } x86emu_t;
diff --git a/src/x86run.c b/src/x86run.c
index bfe832fa..026a145b 100755
--- a/src/x86run.c
+++ b/src/x86run.c
@@ -13,10 +13,13 @@
 
 int Run(x86emu_t *emu)
 {
+    printf_log(LOG_DEBUG, "Run X86, EIP=%p\n", emu, R_EIP);
     emu->quit = 0;
     while (!emu->quit)
     {
-        if(emu->dec) {
+        if(emu->dec && (
+                (emu->trace_end == 0) 
+             || ((R_EIP >= emu->trace_start) && (R_EIP < emu->trace_end))) ) {
             printf_log(LOG_NONE, "%s", DumpCPURegs(emu));
             if(Peek(emu, 0)==0xcc && Peek(emu, 1)=='S' && Peek(emu, 2)=='C') {
                 uint32_t a = *(uint32_t*)(R_EIP+3);
@@ -26,7 +29,7 @@ int Run(x86emu_t *emu)
                     printf_log(LOG_NONE, "%08p: Native call to %p\n", R_EIP, a);
                 }
             } else {
-                printf_log(LOG_NONE, "%08p: %s\n", R_EIP, DecodeX86Trace(emu->dec, R_EIP));
+                printf_log(LOG_NONE, "%s\n", DecodeX86Trace(emu->dec, R_EIP));
             }
         }
         uint8_t opcode = Fetch8(emu);
@@ -242,6 +245,16 @@ int Run(x86emu_t *emu)
                 if(!ACCESS_FLAG(F_ZF))
                     R_EIP += tmp8s;
                 break;
+            case 0x76:  /* JBE Ib */
+                tmp8s = Fetch8s(emu);
+                if((ACCESS_FLAG(F_ZF) || ACCESS_FLAG(F_CF)))
+                    R_EIP += tmp8s;
+                break;
+            case 0x77:  /* JNBE Ib */
+                tmp8s = Fetch8s(emu);
+                if(!(ACCESS_FLAG(F_ZF) || ACCESS_FLAG(F_CF)))
+                    R_EIP += tmp8s;
+                break;
             case 0x7C: /* JL Ib */
                 tmp8s = Fetch8s(emu);
                 if(ACCESS_FLAG(F_SF) != ACCESS_FLAG(F_OF))
@@ -326,6 +339,9 @@ int Run(x86emu_t *emu)
                 GetG(emu, &op2, nextop);
                 op2->dword[0] = (uint32_t)&op1->dword[0];
                 break;
+            
+            case 0x90: /* NOP */
+                break;
 
             case 0xA1: /* MOV EAX, Od */
                 R_EAX = *(uint32_t*)Fetch32(emu);
diff --git a/tests/ref06.txt b/tests/ref06.txt
new file mode 100755
index 00000000..8fca60ff
--- /dev/null
+++ b/tests/ref06.txt
@@ -0,0 +1,4 @@
+[02] Second thread executing
+[02] Thread done.
+
+[00] Done.
diff --git a/tests/test06 b/tests/test06
new file mode 100755
index 00000000..46e3858d
Binary files /dev/null and b/tests/test06 differ
diff --git a/tests/test06.c b/tests/test06.c
new file mode 100755
index 00000000..5a2e0ba0
--- /dev/null
+++ b/tests/test06.c
@@ -0,0 +1,55 @@
+#include <stdio.h>
+#include <string.h>
+#include <pthread.h>
+#include <unistd.h>
+
+const int thread_count = 2;
+pthread_t tid[2];
+const char *thread_messages[2] = {
+	"First thread executing",
+	"Second thread executing"
+};
+
+void *doSomething(void *arg)
+{
+	pthread_t id = pthread_self();
+	int num = -1;
+
+	for (int i = 0 ; i < thread_count ; ++i)
+	{
+		if (pthread_equal(id, tid[i]))
+		{
+			num = i + 1;
+			if (num == 2) printf("[%02d] %s\n", num, thread_messages[i]);
+			break;
+		}
+	}
+
+	for (unsigned int i = 0 ; i < 0x10000 ; ++i);
+	if (num == 2) printf("[%02d] Thread done.\n", num);
+
+	return NULL;
+}
+
+int main(int argc, char const *argv[])
+{
+	int err;
+
+	for (int i = 0 ; i < thread_count ; ++i)
+	{
+		//printf("[00] Thread %d starting\n", i + 1);
+		err = pthread_create(&tid[i], NULL, doSomething, NULL);
+		if (err)
+		{
+			printf("[00] Couldn't create thread %d: %s\n", i + 1, strerror(err));
+		}
+		for (unsigned int i = 0 ; i < 0x1000 ; ++i);
+	}
+
+	//printf("[00] Waiting for all threads to end...\n");
+	for (int i = 0 ; i < thread_count ; ++i)
+		pthread_join(tid[i], NULL);
+	printf("\n[00] Done.\n");
+
+	return 0;
+}