tccrun: exit() via rt_longjmp()

- new LIBTCC API tcc_setjmp() to allow longjmps & signals
  from compiled code back to libtcc per TCCState
- new LIBTCC API tcc_set_backtrace_func() to handle backtrace output
- move c/dtor/atexit stuff to runtime (lib/runmain.c)
- move bt-log.o into libtcc1.a
- add timeouts to github action (beware, it did happen to hang
  infinitely in the signal handler at some point)

16 files changed, 384 insertions, 199 deletions

diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 5fda5d9..e5a8e76 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -7,6 +7,7 @@ on:
 jobs:
   test-x86_64-linux:
     runs-on: ubuntu-20.04
+    timeout-minutes: 2
     steps:
     - uses: actions/checkout@v4
     - name: make & test tcc
@@ -14,6 +15,7 @@ jobs:
 
   test-x86_64-osx:
     runs-on: macos-11
+    timeout-minutes: 2
     steps:
     - uses: actions/checkout@v4
     - name: make & test tcc
@@ -21,6 +23,7 @@ jobs:
 
   test-x86_64-win32:
     runs-on: windows-2019
+    timeout-minutes: 4
     steps:
     - uses: actions/checkout@v4
     - name: make & test tcc
@@ -33,6 +36,7 @@ jobs:
 
   test-armv7-linux:
     runs-on: ubuntu-20.04
+    timeout-minutes: 6
     steps:
     - uses: actions/checkout@v4
     - uses: uraimo/run-on-arch-action@v2
@@ -49,6 +53,7 @@ jobs:
 
   test-aarch64-linux:
     runs-on: ubuntu-20.04
+    timeout-minutes: 6
     steps:
     - uses: actions/checkout@v4
     - uses: uraimo/run-on-arch-action@v2
@@ -65,6 +70,7 @@ jobs:
 
   test-riscv64-linux:
     runs-on: ubuntu-20.04
+    timeout-minutes: 6
     steps:
     - uses: actions/checkout@v4
     - uses: uraimo/run-on-arch-action@v2
diff --git a/Makefile b/Makefile
index 652bf2d..98f01e4 100644
--- a/Makefile
+++ b/Makefile
@@ -29,6 +29,9 @@ ifdef CONFIG_WIN32
  ifneq ($(CONFIG_static),yes)
   LIBTCC = libtcc$(DLLSUF)
   LIBTCCDEF = libtcc.def
+  -LTCC = $(bindir)/libtcc.dll
+ else
+  -LTCC = -ltcc -L$(libdir)
  endif
  ifneq ($(CONFIG_debug),yes)
   LDFLAGS += -s
@@ -70,6 +73,7 @@ else
   endif
   export MACOSX_DEPLOYMENT_TARGET := 10.6
  endif
+ -LTCC = -ltcc
 endif
 
 # run local version of tcc with local libraries and includes
@@ -79,6 +83,15 @@ TCCFLAGS = $(TCCFLAGS$(CFG))
 TCC_LOCAL = $(TOP)/tcc$(EXESUF)
 TCC = $(TCC_LOCAL) $(TCCFLAGS)
 
+# run tests with the installed tcc instead
+ifdef TESTINSTALL
+  TCC_LOCAL = $(bindir)/tcc
+  TCCFLAGS-unx = -I..
+  TCCFLAGS-win = -I.. -B$(bindir)
+  LIBTCC =
+  LIBS += $(-LTCC)
+endif
+
 CFLAGS_P = $(CFLAGS) -pg -static -DCONFIG_TCC_STATIC -DTCC_PROFILE
 LIBS_P = $(LIBS)
 LDFLAGS_P = $(LDFLAGS)
@@ -373,7 +386,7 @@ IR = $(IM) mkdir -p $2 && cp -r $1/. $2
 IM = @echo "-> $2 : $1" ;
 BINCHECK = $(if $(wildcard $(PROGS) *-tcc$(EXESUF)),,@echo "Makefile: nothing found to install" && exit 1)
 
-B_O = bcheck.o bt-exe.o bt-log.o bt-dll.o
+B_O = runmain.o bt-exe.o bt-dll.o bcheck.o
 
 # install progs & libs
 install-unx:
@@ -411,7 +424,7 @@ install-win:
 	$(call IR,$(TOPSRC)/win32/include,"$(tccdir)/include")
 	$(call IR,$(TOPSRC)/win32/examples,"$(tccdir)/examples")
 	$(call IF,$(TOPSRC)/tests/libtcc_test.c,"$(tccdir)/examples")
-	$(call IFw,$(TOPSRC)/libtcc.h libtcc.def,"$(libdir)")
+	$(call IFw,$(TOPSRC)/libtcc.h libtcc.def libtcc.a,"$(libdir)")
 	$(call IFw,$(TOPSRC)/win32/tcc-win32.txt tcc-doc.html,"$(docdir)")
 ifneq "$(wildcard $(LIBTCC1_U))" ""
 	$(call IFw,$(LIBTCC1_U),"$(tccdir)/lib")
@@ -420,9 +433,9 @@ endif
 
 # uninstall on windows
 uninstall-win:
-	@rm -fv $(addprefix "$(bindir)/", libtcc*.dll $(PROGS) *-tcc.exe)
-	@rm -fr $(foreach P,doc examples include lib libtcc,"$(tccdir)/$P/*")
-	@rm -frv $(addprefix "$(tccdir)/", doc examples include lib libtcc)
+	@rm -fv $(foreach P,libtcc*.dll $(PROGS) *-tcc.exe,"$(bindir)"/$P)
+	@rm -fr $(foreach P,doc examples include lib libtcc,"$(tccdir)"/$P/*)
+	@rm -frv $(foreach P,doc examples include lib libtcc,"$(tccdir)"/$P)
 
 # the msys-git shell works to configure && make except it does not have install
 ifeq ($(OS),Windows_NT)
@@ -473,6 +486,9 @@ tcov-tes% : tcc_c$(EXESUF)
 	@$(MAKE) --no-print-directory TCC_LOCAL=$(CURDIR)/$< tes$*
 tcc_c$(EXESUF): $($T_FILES)
 	$S$(TCC) tcc.c -o $@ -ftest-coverage $(DEFINES) $(LIBS)
+# test the installed tcc instead
+test-install: tccdefs_.h
+	@$(MAKE) -C tests TESTINSTALL=yes #_all
 
 clean:
 	@rm -f tcc$(EXESUF) tcc_c$(EXESUF) tcc_p$(EXESUF) *-tcc$(EXESUF)
@@ -506,6 +522,8 @@ help:
 	@echo "   run all/single test(s) from tests/pp"
 	@echo "make tcov-test / tcov-tests2... / tcov-testspp..."
 	@echo "   run tests as above with code coverage. After test(s) see tcc_c$(EXESUF).tcov"
+	@echo "make test-install"
+	@echo "   run tests with the installed tcc"
 	@echo "Other supported make targets:"
 	@echo "   install install-strip doc clean tags ETAGS tar distclean help"
 	@echo "Custom configuration:"
diff --git a/lib/Makefile b/lib/Makefile
index eba3a6b..1a23223 100644
--- a/lib/Makefile
+++ b/lib/Makefile
@@ -67,7 +67,8 @@ OBJ-arm-eabihf = $(ARM_O) $(DSO_O)
 OBJ-arm-wince = $(ARM_O) $(WIN_O)
 OBJ-riscv64 = $(RISCV64_O) $(BCHECK_O) $(DSO_O)
 
-OBJ-extra = $(filter $(B_O),$(OBJ-$T))
+OBJ-extra = $(filter-out bt-log.o,$(filter $(B_O),$(OBJ-$T)))
+OBJ-extra += runmain.o
 OBJ-libtcc1 = $(addprefix $(X),$(filter-out $(OBJ-extra),$(OBJ-$T)))
 
 ALL = $(addprefix $(TOP)/,$(X)libtcc1.a $(OBJ-extra))
diff --git a/lib/bt-exe.c b/lib/bt-exe.c
index 2960dd0..447eaf9 100644
--- a/lib/bt-exe.c
+++ b/lib/bt-exe.c
@@ -26,14 +26,14 @@ void __bt_init(rt_context *p, int is_exe)
 	__bound_init(p->bounds_start, -1);
 
     /* add to chain */
-    WAIT_SEM(&rt_sem);
+    rt_wait_sem();
     p->next = g_rc, g_rc = p;
-    if (is_exe)
+    rt_post_sem();
+    if (is_exe) {
         /* we are the executable (not a dll) */
         p->top_func = main;
-    POST_SEM(&rt_sem);
-    if (is_exe)
         set_exception_handler();
+    }
 }
 
 __declspec(dllexport)
@@ -48,13 +48,13 @@ void __bt_exit(rt_context *p)
 	__bound_exit_dll(p->bounds_start);
 
     /* remove from chain */
-    WAIT_SEM(&rt_sem);
+    rt_wait_sem();
     for (pp = &g_rc; rc = *pp, rc; pp = &rc->next)
         if (rc == p) {
             *pp = rc->next;
             break;
         }
-    POST_SEM(&rt_sem);
+    rt_post_sem();
 }
 
 /* copy a string and truncate it. */
diff --git a/lib/bt-log.c b/lib/bt-log.c
index 05f7c40..87e411d 100644
--- a/lib/bt-log.c
+++ b/lib/bt-log.c
@@ -18,25 +18,24 @@
 #pragma GCC diagnostic ignored "-Wframe-address"
 #endif
 
-typedef struct rt_frame
-{
+typedef struct rt_frame {
     void *ip, *fp, *sp;
 } rt_frame;
 
 __attribute__((weak))
-int _rt_error(rt_frame *f, const char *msg, const char *fmt, va_list ap);
+int __rt_dump(rt_frame *f, const char *msg, const char *fmt, va_list ap);
 
 DLL_EXPORT int tcc_backtrace(const char *fmt, ...)
 {
     va_list ap;
     int ret;
 
-    if (_rt_error) {
+    if (__rt_dump) {
         rt_frame f;
         f.fp = __builtin_frame_address(1);
         f.ip = __builtin_return_address(0);
         va_start(ap, fmt);
-        ret = _rt_error(&f, "", fmt, ap);
+        ret = __rt_dump(&f, NULL, fmt, ap);
         va_end(ap);
     } else {
         const char *p, *nl = "\n";
diff --git a/lib/runmain.c b/lib/runmain.c
new file mode 100644
index 0000000..5ebeeec
--- /dev/null
+++ b/lib/runmain.c
@@ -0,0 +1,87 @@
+/* ------------------------------------------------------------- */
+/* support for tcc_run() */
+
+#ifdef __leading_underscore
+# define _(s) s
+#else
+# define _(s) _##s
+#endif
+
+#ifndef _WIN32
+extern void (*_(_init_array_start)[]) (int argc, char **argv, char **envp);
+extern void (*_(_init_array_end)[]) (int argc, char **argv, char **envp);
+static void run_ctors(int argc, char **argv, char **env)
+{
+    int i = 0;
+    while (&_(_init_array_start)[i] != _(_init_array_end))
+        (*_(_init_array_start)[i++])(argc, argv, env);
+}
+#endif
+
+extern void (*_(_fini_array_start)[]) (void);
+extern void (*_(_fini_array_end)[]) (void);
+static void run_dtors(void)
+{
+    int i = 0;
+    while (&_(_fini_array_end)[i] != _(_fini_array_start))
+        (*_(_fini_array_end)[--i])();
+}
+
+static void *rt_exitfunc[32];
+static void *rt_exitarg[32];
+int __rt_nr_exit;
+
+void __run_on_exit(int ret)
+{
+    int n = __rt_nr_exit;
+    while (n)
+	--n, ((void(*)(int,void*))rt_exitfunc[n])(ret, rt_exitarg[n]);
+}
+
+int on_exit(void *function, void *arg)
+{
+    int n = __rt_nr_exit;
+    if (n < 32) {
+	rt_exitfunc[n] = function;
+	rt_exitarg[n] = arg;
+        __rt_nr_exit = n + 1;
+        return 0;
+    }
+    return 1;
+}
+
+int atexit(void (*function)(void))
+{
+    return on_exit(function, 0);
+}
+
+typedef struct rt_frame {
+    void *ip, *fp, *sp;
+} rt_frame;
+
+void __rt_longjmp(rt_frame *, int);
+
+void exit(int code)
+{
+    rt_frame f;
+    run_dtors();
+    __run_on_exit(code);
+    f.fp = __builtin_frame_address(1);
+    f.ip = __builtin_return_address(0);
+    __rt_longjmp(&f, code);
+}
+
+#ifndef _WIN32
+int main(int, char**, char**);
+
+int _runmain(int argc, char **argv, char **envp)
+{
+    int ret;
+    __rt_nr_exit = 0;
+    run_ctors(argc, argv, envp);
+    ret = main(argc, argv, envp);
+    run_dtors();
+    __run_on_exit(ret);
+    return ret;
+}
+#endif
diff --git a/libtcc.h b/libtcc.h
index 205ed15..7e032d8 100644
--- a/libtcc.h
+++ b/libtcc.h
@@ -100,6 +100,17 @@ LIBTCCAPI void *tcc_get_symbol(TCCState *s, const char *name);
 LIBTCCAPI void tcc_list_symbols(TCCState *s, void *ctx,
     void (*symbol_cb)(void *ctx, const char *name, const void *val));
 
+/* experimental/advanced section (see libtcc_test_mt.c for an example) */
+
+/* catch runtime exceptions (optionally limit backtraces at top_func),
+   when using tcc_set_options("-bt") and when not using tcc_run() */
+LIBTCCAPI void *_tcc_setjmp(TCCState *s1, void *longjmp, void *jmp_buf, void *top_func);
+#define tcc_setjmp(s1,jb,f) setjmp(_tcc_setjmp(s1, longjmp, jb, f))
+
+/* set custom error printer for runtime exceptions */
+typedef void TCCBtFunc(void *pc, const char *file, int line, const char* func);
+LIBTCCAPI void tcc_set_backtrace_func(TCCState *s1, TCCBtFunc*);
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/tcc.h b/tcc.h
index 837dc6e..16f316d 100644
--- a/tcc.h
+++ b/tcc.h
@@ -1000,6 +1000,10 @@ struct TCCState {
 #ifdef _WIN64
     void *run_function_table; /* unwind data */
 #endif
+    struct TCCState *next;
+    struct rt_context *rc; /* pointer to backtrace info block */
+    void *run_lj, *run_jb; /* sj/lj for tcc_setjmp()/tcc_run() */
+    void (*bt_func)(void *, const char*, int, const char*);
 #endif
 
 #ifdef CONFIG_TCC_BACKTRACE
diff --git a/tccelf.c b/tccelf.c
index 2ccf981..049d674 100644
--- a/tccelf.c
+++ b/tccelf.c
@@ -1778,8 +1778,6 @@ ST_FUNC void tcc_add_runtime(TCCState *s1)
         if (s1->do_backtrace) {
             if (s1->output_type & TCC_OUTPUT_EXE)
                 tcc_add_support(s1, "bt-exe.o");
-            if (s1->output_type != TCC_OUTPUT_DLL)
-                tcc_add_support(s1, "bt-log.o");
             tcc_add_btstub(s1);
             lpthread = 1;
         }
diff --git a/tccpe.c b/tccpe.c
index 936f146..d2a2c8b 100644
--- a/tccpe.c
+++ b/tccpe.c
@@ -1961,8 +1961,6 @@ static void pe_add_runtime(TCCState *s1, struct pe_info *pe)
             tcc_add_support(s1, "bt-exe.o");
         if (s1->output_type == TCC_OUTPUT_DLL)
             tcc_add_support(s1, "bt-dll.o");
-        if (s1->output_type != TCC_OUTPUT_DLL)
-            tcc_add_support(s1, "bt-log.o");
         tcc_add_btstub(s1);
     }
 #endif
diff --git a/tccrun.c b/tccrun.c
index 54aa9d6..6ba61a6 100644
--- a/tccrun.c
+++ b/tccrun.c
@@ -26,7 +26,7 @@
 #ifdef CONFIG_TCC_BACKTRACE
 typedef struct rt_context
 {
-    /* --> tccelf.c:tcc_add_btstub wants those below in that order: */
+    /* tccelf.c:tcc_add_btstub() wants these in that order: */
     union {
 	struct {
 	    Stab_Sym *stab_sym;
@@ -42,55 +42,33 @@ typedef struct rt_context
     ElfW(Sym) *esym_start;
     ElfW(Sym) *esym_end;
     char *elf_str;
-
+    // 6
     addr_t prog_base;
     void *bounds_start;
     void *top_func;
     TCCState *s1;
     struct rt_context *next;
-
+    // 11
     int num_callers;
     int dwarf;
-    /* <-- */
 } rt_context; /* size = 11 * PTR_SIZE + 2 * sizeof (int) */
 
-typedef struct rt_frame
-{
+typedef struct rt_frame {
     addr_t ip, fp, sp;
 } rt_frame;
 
 /* linked list of rt_contexts */
 static rt_context *g_rc;
+static TCCState *g_s1;
 /* semaphore to protect it */
 TCC_SEM(static rt_sem);
 static int signal_set;
 static void set_exception_handler(void);
-
-#ifdef _WIN32
-#define	ATTR_WEAK
-#else
-#define	ATTR_WEAK	__attribute__((weak))
-#endif
-int ATTR_WEAK _rt_error(rt_frame *f, const char *msg, const char *fmt, va_list ap);
-void ATTR_WEAK rt_wait_sem(void) { WAIT_SEM(&rt_sem); }
-void ATTR_WEAK rt_post_sem(void) { POST_SEM(&rt_sem); }
-#undef ATTR_WEAK
+static void rt_wait_sem(void) { WAIT_SEM(&rt_sem); }
+static void rt_post_sem(void) { POST_SEM(&rt_sem); }
+static int rt_get_caller_pc(addr_t *paddr, rt_frame *f, int level);
 #endif /* def CONFIG_TCC_BACKTRACE */
 
-/* handle exit/atexit for tcc_run() -- thread-unsafe */
-static jmp_buf rt_jb;
-static int rt_do_jmp;
-static int rt_nr_exit;
-static void *rt_exitfunc[32];
-static void *rt_exitarg[32];
-
-static void rt_exit(int code)
-{
-    if (rt_do_jmp)
-        longjmp(rt_jb, code ? code : 256);
-    exit(code);
-}
-
 /* ------------------------------------------------------------- */
 /* defined when included from lib/bt-exe.c */
 #ifndef CONFIG_TCC_BACKTRACE_ONLY
@@ -101,7 +79,8 @@ static void rt_exit(int code)
 
 static int protect_pages(void *ptr, unsigned long length, int mode);
 static int tcc_relocate_ex(TCCState *s1, void *ptr, unsigned ptr_diff);
-
+static int __rt_dump(rt_frame *f, const char *msg, const char *fmt, va_list ap);
+static void rt_longjmp(rt_frame *f, int code);
 #ifdef _WIN64
 static void *win64_add_function_table(TCCState *s1);
 static void win64_del_function_table(void *);
@@ -118,12 +97,9 @@ static void bt_link(TCCState *s1)
     rc = tcc_get_symbol(s1, "__rt_info");
     if (!rc)
         return;
-    rt_wait_sem();
-    rc->next = g_rc, g_rc = rc, rc->s1 = s1;
     rc->esym_start = (ElfW(Sym) *)(symtab_section->data);
     rc->esym_end = (ElfW(Sym) *)(symtab_section->data + symtab_section->data_offset);
     rc->elf_str = (char *)symtab_section->link->data;
-    rc->top_func = tcc_get_symbol(s1, "main");
     if (PTR_SIZE == 8 && !s1->dwarf)
         rc->prog_base &= 0xffffffff00000000ULL;
 #ifdef CONFIG_TCC_BCHECK
@@ -132,7 +108,7 @@ static void bt_link(TCCState *s1)
             ((void(*)(void*,int))p)(rc->bounds_start, 1);
     }
 #endif
-    rt_post_sem();
+    rc->next = g_rc, g_rc = rc, rc->s1 = s1, s1->rc = rc;
     if (0 == signal_set)
         set_exception_handler(), signal_set = 1;
 #endif
@@ -142,16 +118,35 @@ static void bt_unlink(TCCState *s1)
 {
 #ifdef CONFIG_TCC_BACKTRACE
     rt_context *rc, **pp;
-    rt_wait_sem();
     for (pp = &g_rc; rc = *pp, rc; pp = &rc->next)
         if (rc->s1 == s1) {
             *pp = rc->next;
             break;
         }
-    rt_post_sem();
 #endif
 }
 
+static void st_link(TCCState *s1)
+{
+    rt_wait_sem();
+    s1->next = g_s1, g_s1 = s1;
+    bt_link(s1);
+    rt_post_sem();
+}
+
+static void st_unlink(TCCState *s1)
+{
+    TCCState *s2, **pp;
+    rt_wait_sem();
+    bt_unlink(s1);
+    for (pp = &g_s1; s2 = *pp, s2; pp = &s2->next)
+        if (s2 == s1) {
+            *pp = s2->next;
+            break;
+        }
+    rt_post_sem();
+}
+
 #if !_WIN32 && !__APPLE__
 //#define HAVE_SELINUX 1
 #endif
@@ -196,9 +191,8 @@ LIBTCCAPI int tcc_relocate(TCCState *s1)
         exit(tcc_error_noabort("'tcc_relocate()' twice is no longer supported"));
 
 #ifdef CONFIG_TCC_BACKTRACE
-    /* for bt-log.c (but not when 'tcc -bt -run tcc.c') */
-    if (s1->do_backtrace && !tcc_get_symbol(s1, "_rt_error"))
-        tcc_add_symbol(s1, "_rt_error", _rt_error);
+    if (s1->do_backtrace)
+        tcc_add_symbol(s1, "__rt_dump", __rt_dump); /* for bt-log.c */
 #endif
 
     size = tcc_relocate_ex(s1, NULL, 0);
@@ -209,7 +203,7 @@ LIBTCCAPI int tcc_relocate(TCCState *s1)
         return -1;
     ret = tcc_relocate_ex(s1, s1->run_ptr, ptr_diff);
     if (ret == 0)
-        bt_link(s1);
+        st_link(s1);
     return ret;
 }
 
@@ -231,12 +225,11 @@ ST_FUNC void tcc_run_free(TCCState *s1)
     }
     /* free loaded dlls array */
     dynarray_reset(&s1->loaded_dlls, &s1->nb_loaded_dlls);
-
     /* unmap or unprotect and free memory */
     ptr = s1->run_ptr;
     if (NULL == ptr)
         return;
-    bt_unlink(s1);
+    st_unlink(s1);
     size = s1->run_size;
 #ifdef HAVE_SELINUX
     munmap(ptr, size * 2);
@@ -250,41 +243,25 @@ ST_FUNC void tcc_run_free(TCCState *s1)
 #endif
 }
 
-static void run_cdtors(TCCState *s1, const char *start, const char *end,
-                       int argc, char **argv, char **envp)
+LIBTCCAPI void *_tcc_setjmp(TCCState *s1, void *p_longjmp, void *p_jmp_buf, void *func)
 {
-    void **a = (void **)get_sym_addr(s1, start, 0, 0);
-    void **b = (void **)get_sym_addr(s1, end, 0, 0);
-    while (a != b)
-        ((void(*)(int, char **, char **))*a++)(argc, argv, envp);
+    s1->run_lj = p_longjmp;
+    s1->run_jb = p_jmp_buf;
+    if (func && s1->rc)
+        s1->rc->top_func = func;
+    return p_jmp_buf;
 }
 
-static void run_on_exit(int ret)
+LIBTCCAPI void tcc_set_backtrace_func(TCCState *s1, TCCBtFunc *func)
 {
-    int n = rt_nr_exit;
-    while (n)
-	--n, ((void(*)(int,void*))rt_exitfunc[n])(ret, rt_exitarg[n]);
-}
-
-static int rt_on_exit(void *function, void *arg)
-{
-    if (rt_nr_exit < countof(rt_exitfunc)) {
-	rt_exitfunc[rt_nr_exit] = function;
-	rt_exitarg[rt_nr_exit++] = arg;
-        return 0;
-    }
-    return 1;
-}
-
-static int rt_atexit(void *function)
-{
-    return rt_on_exit(function, NULL);
+    s1->bt_func = func;
 }
 
 /* launch the compiled program with the given arguments */
 LIBTCCAPI int tcc_run(TCCState *s1, int argc, char **argv)
 {
     int (*prog_main)(int, char **, char **), ret;
+    jmp_buf main_jb;
 
 #if defined(__APPLE__) || defined(__FreeBSD__)
     char **envp = NULL;
@@ -295,35 +272,29 @@ LIBTCCAPI int tcc_run(TCCState *s1, int argc, char **argv)
     char **envp = environ;
 #endif
 
-    s1->run_main = s1->nostdlib ? "_start" : "main";
-    if ((s1->dflag & 16) && (addr_t)-1 == get_sym_addr(s1, s1->run_main, 0, 1))
+    /* tcc -dt -run ... nothing to do if no main() */
+    if ((s1->dflag & 16) && (addr_t)-1 == get_sym_addr(s1, "main", 0, 1))
         return 0;
 
-    tcc_add_symbol(s1, "exit", rt_exit);
-    tcc_add_symbol(s1, "atexit", rt_atexit);
-    tcc_add_symbol(s1, "on_exit", rt_on_exit);
+    tcc_add_support(s1, "runmain.o");
+    tcc_add_symbol(s1, "__rt_longjmp", rt_longjmp);
+    s1->run_main = (s1->nostdlib ? "_start" : "_runmain");
     if (tcc_relocate(s1) < 0)
         return -1;
 
     prog_main = (void*)get_sym_addr(s1, s1->run_main, 1, 1);
     if ((addr_t)-1 == (addr_t)prog_main)
         return -1;
-
     errno = 0; /* clean errno value */
     fflush(stdout);
     fflush(stderr);
-    rt_do_jmp = 1;
-    rt_nr_exit = 0;
 
-    /* These aren't C symbols, so don't need leading underscore handling.  */
-    run_cdtors(s1, "__init_array_start", "__init_array_end", argc, argv, envp);
-    ret = setjmp(rt_jb);
+    ret = tcc_setjmp(s1, main_jb, tcc_get_symbol(s1, "main"));
     if (0 == ret)
         ret = prog_main(argc, argv, envp);
     else if (256 == ret)
         ret = 0;
-    run_cdtors(s1, "__fini_array_start", "__fini_array_end", 0, NULL, NULL);
-    run_on_exit(ret);
+
     if (s1->dflag & 16 && ret) /* tcc -dt -run ... */
         fprintf(s1->ppfp, "[returns %d]\n", ret), fflush(s1->ppfp);
     return ret;
@@ -385,7 +356,6 @@ static int tcc_relocate_ex(TCCState *s1, void *ptr, unsigned ptr_diff)
     addr_t mem, addr;
 
     if (NULL == ptr) {
-        s1->nb_errors = 0;
 #ifdef TCC_TARGET_PE
         pe_output_file(s1, NULL);
 #else
@@ -616,7 +586,7 @@ static char *rt_elfsym(rt_context *rc, addr_t wanted_pc, addr_t *func_addr)
 /* print the position in the source file of PC value 'pc' by reading
    the stabs debug information */
 static addr_t rt_printline (rt_context *rc, addr_t wanted_pc,
-    const char *msg, const char *skip)
+    rt_context** prc, const char *msg, const char *skip)
 {
     char func_name[128];
     addr_t func_addr, last_pc, pc;
@@ -726,14 +696,21 @@ next:
 
 found:
     i = last_incl_index;
+    str = NULL;
     if (i > 0) {
         str = incl_files[--i];
         if (skip[0] && strstr(str, skip))
             return (addr_t)-1;
-        rt_printf("%s:%d: ", str, last_line_num);
-    } else
-        rt_printf("%08llx : ", (long long)wanted_pc);
-    rt_printf("%s %s", msg, func_name[0] ? func_name : "???");
+    }
+    if (rc && rc->s1 && rc->s1->bt_func) {
+        rc->s1->bt_func((void*)wanted_pc, str, last_line_num, func_name[0] ? func_name : NULL);
+    } else {
+        if (str)
+            rt_printf("%s:%d: ", str, last_line_num);
+        else
+            rt_printf("%08llx : ", (long long)wanted_pc);
+        rt_printf("%s %s", msg, func_name[0] ? func_name : "???");
+    }
 #if 0
     if (--i >= 0) {
         rt_printf(" (included from ");
@@ -746,6 +723,7 @@ found:
         rt_printf(")");
     }
 #endif
+    *prc = rc;
     return func_addr;
 }
 
@@ -816,7 +794,7 @@ dwarf_read_sleb128(unsigned char **ln, unsigned char *end)
 }
 
 static addr_t rt_printline_dwarf (rt_context *rc, addr_t wanted_pc,
-    const char *msg, const char *skip)
+    rt_context** prc, const char *msg, const char *skip)
 {
     unsigned char *ln;
     unsigned char *cp;
@@ -859,6 +837,7 @@ static addr_t rt_printline_dwarf (rt_context *rc, addr_t wanted_pc,
 next:
     filename = NULL;
     func_addr = 0;
+    line = 0;
     if (NULL == rc)
         goto found;
 
@@ -1102,29 +1081,33 @@ next_line:
     goto next;
 
 found:
-    if (filename) {
-	if (skip[0] && strstr(filename, skip))
-	    return (addr_t)-1;
-	rt_printf("%s:%d: ", filename, line);
+    if (rc && rc->s1 && rc->s1->bt_func) {
+        rc->s1->bt_func((void*)wanted_pc, filename, line, function);
+    } else {
+        if (filename) {
+            if (skip[0] && strstr(filename, skip))
+                return (addr_t)-1;
+            rt_printf("%s:%d: ", filename, line);
+        }
+        else
+            rt_printf("0x%08llx : ", (long long)wanted_pc);
+        rt_printf("%s %s", msg, function ? function : "???");
     }
-    else
-	rt_printf("0x%08llx : ", (long long)wanted_pc);
-    rt_printf("%s %s", msg, function ? function : "???");
+    *prc = rc;
     return (addr_t)func_addr;
 }
 /* ------------------------------------------------------------- */
-
-static int rt_get_caller_pc(addr_t *paddr, rt_frame *f, int level);
-
-int _rt_error(rt_frame *f, const char *msg, const char *fmt, va_list ap)
+#ifndef CONFIG_TCC_BACKTRACE_ONLY
+static
+#endif
+int __rt_dump(rt_frame *f, const char *msg, const char *fmt, va_list ap)
 {
-    rt_context *rc;
+    rt_context *rc, *rd;
     addr_t pc = 0;
     char skip[100];
     int i, level, ret, n, one;
     const char *a, *b;
-    addr_t (*printline)(rt_context*, addr_t, const char*, const char*);
-    addr_t top_func = 0;
+    addr_t (*printline)(rt_context*, addr_t, rt_context**, const char*, const char*);
 
     skip[0] = 0;
     /* If fmt is like "^file.c^..." then skip calls from 'file.c' */
@@ -1145,27 +1128,28 @@ int _rt_error(rt_frame *f, const char *msg, const char *fmt, va_list ap)
             printline = rt_printline_dwarf;
         if (rc->num_callers)
             n = rc->num_callers;
-        top_func = (addr_t)rc->top_func;
     }
-
     for (i = level = 0; level < n; i++) {
         ret = rt_get_caller_pc(&pc, f, i);
-        a = "%s";
         if (ret != -1) {
-            pc = printline(rc, pc, level ? "by" : "at", skip);
+            pc = printline(rc, pc, &rd, level ? "by" : "at", skip);
             if (pc == (addr_t)-1)
                 continue;
-            a = ": %s";
         }
         if (level == 0) {
-            rt_printf(a, msg);
+            if (rd && rd->s1 && rd->s1->bt_func)
+                break;
+            if (ret != -1)
+                rt_printf(": ");
+            if (msg)
+                rt_printf("%s: ", msg);
             rt_vprintf(fmt, ap);
         } else if (ret == -1)
             break;
         if (one)
             break;
         rt_printf("\n");
-        if (ret == -1 || (pc == top_func && pc))
+        if (ret == -1 || (rd && pc == (addr_t)rd->top_func && pc))
             break;
         ++level;
     }
@@ -1180,11 +1164,46 @@ static int rt_error(rt_frame *f, const char *fmt, ...)
     va_list ap;
     int ret;
     va_start(ap, fmt);
-    ret = _rt_error(f, "RUNTIME ERROR: ", fmt, ap);
+    ret = __rt_dump(f, "RUNTIME ERROR", fmt, ap);
     va_end(ap);
     return ret;
 }
 
+static TCCState *rt_find_state(rt_frame *f)
+{
+    TCCState *s;
+    int level;
+    addr_t pc;
+
+    rt_wait_sem();
+    for (s = g_s1; s; s = s->next) {
+        if (0 == s->run_lj)
+            continue;
+        for (level = 0; level < 8; ++level) {
+            if (rt_get_caller_pc(&pc, f, level) < 0)
+                break;
+            if (pc >= (addr_t)s->run_ptr
+             && pc < (addr_t)s->run_ptr + s->run_size)
+                goto found;
+        }
+    }
+found:
+    rt_post_sem();
+    //fprintf(stderr, "\nrt_state found %s %p %p\n", s ? "YES" : "NO", s, s->rc->top_func), fflush(stderr);
+    return s;
+}
+
+static void rt_longjmp(rt_frame *f, int code)
+{
+    TCCState *s = rt_find_state(f);
+    if (s && s->run_lj) {
+        if (code == 0)
+            code = 256;
+        ((void(*)(void*,int))s->run_lj)(s->run_jb, code);
+    }
+    exit(code);
+}
+
 /* ------------------------------------------------------------- */
 
 #ifndef _WIN32
@@ -1319,7 +1338,8 @@ static void sig_error(int signum, siginfo_t *siginf, void *puc)
         rt_error(&f, "caught signal %d", signum);
         break;
     }
-    rt_exit(255);
+    set_exception_handler();
+    rt_longjmp(&f, 255);
 }
 
 #ifndef SA_SIGINFO
@@ -1333,7 +1353,7 @@ static void set_exception_handler(void)
     /* install TCC signal handlers to print debug info on fatal
        runtime errors */
     sigemptyset (&sigact.sa_mask);
-    sigact.sa_flags = SA_SIGINFO | SA_RESETHAND;
+    sigact.sa_flags = SA_SIGINFO | SA_RESETHAND | SA_NODEFER;
 #if 0//def SIGSTKSZ // this causes signals not to work at all on some (older) linuxes
     sigact.sa_flags |= SA_ONSTACK;
 #endif
@@ -1365,7 +1385,6 @@ static long __stdcall cpu_exception_handler(EXCEPTION_POINTERS *ex_info)
 {
     rt_frame f;
     unsigned code;
-
     rt_getcontext(ex_info->ContextRecord, &f);
 
     switch (code = ex_info->ExceptionRecord->ExceptionCode) {
@@ -1387,8 +1406,7 @@ static long __stdcall cpu_exception_handler(EXCEPTION_POINTERS *ex_info)
         rt_error(&f, "caught exception %08x", code);
         break;
     }
-    if (rt_do_jmp)
-        rt_exit(255);
+    rt_longjmp(&f, 255);
     return EXCEPTION_EXECUTE_HANDLER;
 }
 
@@ -1405,45 +1423,41 @@ static void set_exception_handler(void)
 #if defined(__i386__) || defined(__x86_64__)
 static int rt_get_caller_pc(addr_t *paddr, rt_frame *rc, int level)
 {
-    addr_t ip, fp;
     if (level == 0) {
-        ip = rc->ip;
+        *paddr = rc->ip;
     } else {
-        ip = 0;
-        fp = rc->fp;
-        while (--level) {
-            /* XXX: check address validity with program info */
-            if (fp <= 0x1000)
+        addr_t fp = rc->fp;
+        while (1) {
+            if (fp < 0x1000)
+                return -1;
+            if (0 == --level)
                 break;
+            /* XXX: check address validity with program info */
             fp = ((addr_t *)fp)[0];
         }
-        if (fp > 0x1000)
-            ip = ((addr_t *)fp)[1];
+        *paddr = ((addr_t *)fp)[1];
     }
-    if (ip <= 0x1000)
-        return -1;
-    *paddr = ip;
     return 0;
 }
 
-#elif defined(__arm__)
+/* XXX: only supports linux/bsd */
+#elif defined(__arm__) && !defined(_WIN32)
 static int rt_get_caller_pc(addr_t *paddr, rt_frame *rc, int level)
 {
-    /* XXX: only supports linux/bsd */
-#if !defined(__linux__) && \
-    !defined(__FreeBSD__) && !defined(__OpenBSD__) && !defined(__NetBSD__)
-    return -1;
-#else
     if (level == 0) {
         *paddr = rc->ip;
     } else {
         addr_t fp = rc->fp;
-        while (--level)
+        while (1) {
+            if (fp < 0x1000)
+                return -1;
+            if (0 == --level)
+                break;
             fp = ((addr_t *)fp)[0];
+        }
         *paddr = ((addr_t *)fp)[2];
     }
     return 0;
-#endif
 }
 
 #elif defined(__aarch64__)
@@ -1452,10 +1466,15 @@ static int rt_get_caller_pc(addr_t *paddr, rt_frame *rc, int level)
     if (level == 0) {
         *paddr = rc->ip;
     } else {
-        addr_t *fp = (addr_t*)rc->fp;
-        while (--level)
-            fp = (addr_t *)fp[0];
-        *paddr = fp[1];
+        addr_t fp = rc->fp;
+        while (1) {
+            if (fp < 0x1000)
+                return -1;
+            if (0 == --level)
+                break;
+            fp = ((addr_t *)fp)[0];
+        }
+        *paddr = ((addr_t *)fp)[1];
     }
     return 0;
 }
@@ -1466,12 +1485,15 @@ static int rt_get_caller_pc(addr_t *paddr, rt_frame *rc, int level)
     if (level == 0) {
         *paddr = rc->ip;
     } else {
-        addr_t *fp = (addr_t*)rc->fp;
-        while (--level && fp >= (addr_t*)0x1000)
-            fp = (addr_t *)fp[-2];
-        if (fp < (addr_t*)0x1000)
-          return -1;
-        *paddr = fp[-1];
+        addr_t fp = rc->fp;
+        while (1) {
+            if (fp < 0x1000)
+                return -1;
+            if (0 == --level)
+                break;
+            fp = ((addr_t *)fp)[-2];
+        }
+        *paddr = ((addr_t *)fp)[-1];
     }
     return 0;
 }
diff --git a/tests/Makefile b/tests/Makefile
index 2d6f64f..54f7c8b 100644
--- a/tests/Makefile
+++ b/tests/Makefile
@@ -14,23 +14,20 @@ TESTS = \
  libtest \
  libtest_mt \
  test3 \
- memtest \
- dlltest \
  abitest \
  asm-c-connect-test \
  vla_test-run \
- cross-test \
  tests2-dir \
- pp-dir
+ pp-dir \
+ memtest \
+ dlltest \
+ cross-test
 
 # test4_static -- Not all relocation types are implemented yet.
 # asmtest / asmtest2 -- minor differences with gcc
 
 ifneq ($(CONFIG_bcheck),no)
- TESTS += btest test1b
- ifndef CONFIG_WIN32
-  TESTS += tccb
- endif
+ TESTS += btest test1b tccb
 endif
 ifeq ($(CONFIG_dll),no)
  TESTS := $(filter-out dlltest, $(TESTS))
@@ -82,7 +79,9 @@ all test :
 	@echo ------------ version ------------
 	@$(TCC_LOCAL) -v
 	@$(MAKE) --no-print-directory -s clean
-	@$(MAKE) --no-print-directory -s -r $(TESTS)
+	@$(MAKE) --no-print-directory -s -r _all
+
+_all : $(TESTS)
 
 hello-exe: ../examples/ex1.c
 	@echo ------------ $@ ------------
@@ -199,14 +198,15 @@ btest: boundtest.c
 	       echo "Test $$i failed as expected" ; \
 	   fi ;\
 	done ;\
-	echo Bound test OK
+	echo Bound-Test OK
 
 tccb:
 	@echo ------------ $@ ------------
-	$(TCC) -b $(TOPSRC)/tcc.c $(TCCFLAGS) $(NATIVE_DEFINES) -o v1-tcc$(EXESUF)
-	mv v1-tcc$(EXESUF) v2-tcc$(EXESUF)
-	./v2-tcc$(EXESUF) -b $(TOPSRC)/tcc.c $(TCCFLAGS) $(NATIVE_DEFINES) -o v1-tcc$(EXESUF)
-	cmp -s v1-tcc$(EXESUF) v2-tcc$(EXESUF) && echo "Bound-Test tcc OK"
+	$(TCC) -b $(TOPSRC)/tcc.c $(TCCFLAGS) $(NATIVE_DEFINES) -o tccb1.exe
+	mv tccb1.exe tccb2.exe
+	./tccb2.exe -b $(TOPSRC)/tcc.c $(TCCFLAGS) $(NATIVE_DEFINES) -o tccb1.exe
+	cmp -s tccb1.exe tccb2.exe && echo "Exe Bound-Rest OK"
+
 
 # speed test
 speedtest: ex2 ex3
diff --git a/tests/libtcc_test_mt.c b/tests/libtcc_test_mt.c
index 8af4e64..6c712c6 100644
--- a/tests/libtcc_test_mt.c
+++ b/tests/libtcc_test_mt.c
@@ -6,6 +6,7 @@
 #include <stdlib.h>
 #include <stdio.h>
 #include <string.h>
+#include <setjmp.h>
 #include "libtcc.h"
 
 #define M 20 /* number of states */
@@ -84,6 +85,8 @@ PROG(my_program)
 "\n"
 "int foo(int n)\n"
 "{\n"
+"    if (n >= N_CRASH && n < N_CRASH + 3)\n"
+"       *(void**)0 = 0;\n"
 "    printf(\" %d\", fib(n));\n"
 "    return 0;\n"
 "#  warning is this the correct file:line...\n"
@@ -110,6 +113,11 @@ void parse_args(TCCState *s)
     }
 }
 
+void bt_func(void *pc, const char *file, int line, const char *func)
+{
+    printf("  *** at %s:%d in '%s'\n", file, line, func);
+}
+
 TCCState *new_state(int w)
 {
     TCCState *s = tcc_new();
@@ -119,7 +127,14 @@ TCCState *new_state(int w)
     }
     tcc_set_error_func(s, stdout, handle_error);
     parse_args(s);
-    if (!w) tcc_set_options(s, "-w");
+    if (0 == (w & 1))
+        tcc_set_options(s, "-w");
+    if (w & 2) {
+        tcc_set_options(s, "-bt");
+        tcc_define_symbol(s, "N_CRASH", str(M/2));
+        tcc_set_backtrace_func(s, bt_func);
+    } else
+        tcc_define_symbol(s, "N_CRASH", "99");
     tcc_set_output_type(s, TCC_OUTPUT_MEMORY);
     return s;
 }
@@ -139,23 +154,28 @@ void *reloc_state(TCCState *s, const char *entry)
 }
 
 /* work with several states at the same time */
-int state_test(void)
+int state_test(int w)
 {
     TCCState *s[M];
-    int (*func[M])(int);
+    int (*funcs[M])(int);
     int n;
+    jmp_buf jb;
 
     for (n = 0; n < M + 4; ++n) {
         unsigned a = n, b = n - 1, c = n - 2, d = n - 3, e = n - 4;
         if (a < M)
-            s[a] = new_state(0);
+            s[a] = new_state(w);
         if (b < M)
             if (tcc_compile_string(s[b], my_program) == -1)
                 break;
         if (c < M)
-            func[c] = reloc_state(s[c], "foo");
-        if (d < M && func[d])
-            func[d](F(d));
+            funcs[c] = reloc_state(s[c], "foo");
+        if (d < M && funcs[d]) {
+            if ((w & 2) && d == 8)
+                printf("\n");
+            if (0 == tcc_setjmp(s[d], jb, funcs[d]))
+                funcs[d](F(d));
+        }
         if (e < M)
             tcc_delete(s[e]);
     }
@@ -257,7 +277,13 @@ int main(int argc, char **argv)
 #if 1
     printf("running fib with mixed calls\n "), fflush(stdout);
     t = getclock_ms();
-    state_test();
+    state_test(0);
+    printf("\n (%u ms)\n", getclock_ms() - t);
+#endif
+#if 1
+    printf("producing some exceptions\n "), fflush(stdout);
+    t = getclock_ms();
+    state_test(2);
     printf("\n (%u ms)\n", getclock_ms() - t);
 #endif
 #if 1
diff --git a/tests/tests2/128_run_atexit.c b/tests/tests2/128_run_atexit.c
index 0748c86..3c7cfbb 100644
--- a/tests/tests2/128_run_atexit.c
+++ b/tests/tests2/128_run_atexit.c
@@ -4,6 +4,11 @@ int atexit(void (*function)(void));
 int on_exit(void (*function)(int, void *), void *arg);
 void exit(int status);
 
+void __attribute((constructor)) startup5(void)
+{
+    printf ("startup5\n");
+}
+
 void cleanup1(void)
 {
     printf ("cleanup1\n");
diff --git a/tests/tests2/128_run_atexit.expect b/tests/tests2/128_run_atexit.expect
index 3305785..fda764c 100644
--- a/tests/tests2/128_run_atexit.expect
+++ b/tests/tests2/128_run_atexit.expect
@@ -1,4 +1,5 @@
 [test_128_return]
+startup5
 cleanup5
 1 cleanup4
 1 cleanup3
@@ -7,6 +8,7 @@ cleanup1
 [returns 1]
 
 [test_128_exit]
+startup5
 cleanup5
 2 cleanup4
 2 cleanup3
diff --git a/win32/lib/crt1.c b/win32/lib/crt1.c
index 0686302..6406ed3 100644
--- a/win32/lib/crt1.c
+++ b/win32/lib/crt1.c
@@ -37,15 +37,6 @@ extern int _tmain(int argc, _TCHAR * argv[], _TCHAR * env[]);
 
 #include "crtinit.c"
 
-static int do_main (int argc, _TCHAR * argv[], _TCHAR * env[])
-{
-    int retval;
-    run_ctors(argc, argv, env);
-    retval = _tmain(__argc, __targv, _tenviron);
-    run_dtors();
-    return retval;
-}
-
 /* Allow command-line globbing with "int _dowildcard = 1;" in the user source */
 int _dowildcard;
 
@@ -56,6 +47,8 @@ static LONG WINAPI catch_sig(EXCEPTION_POINTERS *ex)
 
 void _tstart(void)
 {
+    int ret;
+
     _startupinfo start_info = {0};
     SetUnhandledExceptionFilter(catch_sig);
     // Sets the current application type
@@ -68,11 +61,21 @@ void _tstart(void)
 #endif
 
     __tgetmainargs( &__argc, &__targv, &_tenviron, _dowildcard, &start_info);
-    exit(do_main(__argc, __targv, _tenviron));
+    run_ctors(__argc, __targv, _tenviron);
+    ret = _tmain(__argc, __targv, _tenviron);
+    run_dtors();
+    exit(ret);
 }
 
+// =============================================
+// for 'tcc -run ,,,'
+
+__attribute__((weak)) extern int __rt_nr_exit;
+__attribute__((weak)) extern int __run_on_exit();
+
 int _runtmain(int argc, /* as tcc passed in */ char **argv)
 {
+    int ret;
 #ifdef UNICODE
     _startupinfo start_info = {0};
 
@@ -89,7 +92,12 @@ int _runtmain(int argc, /* as tcc passed in */ char **argv)
 #if defined __i386__ || defined __x86_64__
     _controlfp(_PC_53, _MCW_PC);
 #endif
-    return _tmain(__argc, __targv, _tenviron);
+    __rt_nr_exit = 0;
+    run_ctors(__argc, __targv, _tenviron);
+    ret = _tmain(__argc, __targv, _tenviron);
+    run_dtors();
+    __run_on_exit(ret);
+    return ret;
 }
 
 // =============================================