diff --git a/libtcc.c b/libtcc.c
index 5f27d2e4..0ffd6162 100644
--- a/libtcc.c
+++ b/libtcc.c
@@ -1028,6 +1028,11 @@ LIBTCCAPI void tcc_delete(TCCState *s1)
     dynarray_reset(&s1->target_deps, &s1->nb_target_deps);
     dynarray_reset(&s1->pragma_libs, &s1->nb_pragma_libs);
     dynarray_reset(&s1->argv, &s1->argc);
+#ifdef __OpenBSD__
+    tcc_free(s1->dlopens);
+    s1->dlopens = NULL;
+    s1->nb_dlopens = 0;
+#endif
 
     cstr_free(&s1->cmdline_defs);
     cstr_free(&s1->cmdline_incl);
@@ -1113,7 +1118,10 @@ LIBTCCAPI int tcc_set_output_type(TCCState *s, int output_type)
 #if defined(__OpenBSD__)
         if (output_type != TCC_OUTPUT_DLL)
 	    tcc_add_crt(s, "crt0.o");
-        tcc_add_crt(s, "crtbegin.o");
+        if (output_type == TCC_OUTPUT_DLL)
+            tcc_add_crt(s, "crtbeginS.o");
+        else
+            tcc_add_crt(s, "crtbegin.o");
 #elif defined(__FreeBSD__)
         if (output_type != TCC_OUTPUT_DLL)
             tcc_add_crt(s, "crt1.o");
@@ -1181,8 +1189,15 @@ ST_FUNC int tcc_add_file_internal(TCCState *s1, const char *filename, int flags)
             if (s1->output_type == TCC_OUTPUT_MEMORY) {
                 ret = 0;
 #ifdef TCC_IS_NATIVE
-                if (NULL == dlopen(filename, RTLD_GLOBAL | RTLD_LAZY))
-                    ret = -1;
+		{
+                    void *dl = dlopen(filename, RTLD_GLOBAL | RTLD_LAZY);
+                    if (NULL == dl)
+                        ret = -1;
+#ifdef __OpenBSD__
+		    else
+			dynarray_add(&s1->dlopens, &s1->nb_dlopens, dl);
+#endif
+		}
 #endif
             } else {
 #ifndef TCC_TARGET_MACHO
diff --git a/tcc.h b/tcc.h
index 34ce5cd4..31b342b8 100644
--- a/tcc.h
+++ b/tcc.h
@@ -793,6 +793,12 @@ struct TCCState {
     char **crt_paths;
     int nb_crt_paths;
 
+#ifdef __OpenBSD__
+    /* track dlopen */
+    void **dlopens;
+    int nb_dlopens;
+#endif
+
     /* -D / -U options */
     CString cmdline_defs;
     /* -include options */
diff --git a/tccelf.c b/tccelf.c
index 0a569da8..856b8dcc 100644
--- a/tccelf.c
+++ b/tccelf.c
@@ -912,6 +912,14 @@ ST_FUNC void relocate_syms(TCCState *s1, Section *symtab, int do_resolve)
                 void *addr = dlsym(RTLD_DEFAULT, name + 1);
 #else
                 void *addr = dlsym(RTLD_DEFAULT, name);
+#ifdef __OpenBSD__
+		if (addr == NULL) {
+		    int i;
+		    for (i = 0; i < s1->nb_dlopens; i++)
+                        if ((addr = dlsym(s1->dlopens[i], name)))
+			    break;
+		}
+#endif
 #endif
                 if (addr) {
                     sym->st_value = (addr_t) addr;
@@ -1413,7 +1421,7 @@ ST_FUNC void tcc_add_runtime(TCCState *s1)
 #ifdef CONFIG_TCC_BCHECK
         if (s1->do_bounds_check && s1->output_type != TCC_OUTPUT_DLL) {
             tcc_add_library_err(s1, "pthread");
-#if !defined(__OpenBSD__)
+#if !defined(__OpenBSD__) && !defined(__NetBSD__)
             tcc_add_library_err(s1, "dl");
 #endif
             tcc_add_support(s1, "bcheck.o");
@@ -1433,7 +1441,9 @@ ST_FUNC void tcc_add_runtime(TCCState *s1)
             tcc_add_support(s1, TCC_LIBTCC1);
 #if defined(__OpenBSD__) || defined(__FreeBSD__) || defined(__NetBSD__)
         /* add crt end if not memory output */
-	if (s1->output_type != TCC_OUTPUT_MEMORY) {
+	if (s1->output_type == TCC_OUTPUT_DLL)
+	    tcc_add_crt(s1, "crtendS.o");
+	else if (s1->output_type != TCC_OUTPUT_MEMORY) {
 	    tcc_add_crt(s1, "crtend.o");
 #if defined(__FreeBSD__) || defined(__NetBSD__)
             tcc_add_crt(s1, "crtn.o");
diff --git a/tccpp.c b/tccpp.c
index 992019fa..49c4d39f 100644
--- a/tccpp.c
+++ b/tccpp.c
@@ -3700,7 +3700,8 @@ static void tcc_predefs(CString *cstr)
     "__BOTH(char*,strcat,(char*,const char*))\n"
     "__BOTH(char*,strchr,(const char*,int))\n"
     "__BOTH(char*,strdup,(const char*))\n"
-#ifdef TCC_TARGET_PE
+#if defined(TCC_TARGET_PE) || defined(__OpenBSD__) || \
+    defined(__FreeBSD__) || defined(__NetBSD__)
     "#define __MAYBE_REDIR __BOTH\n"
 #else  // HAVE MALLOC_REDIR
     "#define __MAYBE_REDIR __BUILTIN\n"
diff --git a/tests/asm-c-connect-1.c b/tests/asm-c-connect-1.c
index d79270d4..90f59c83 100644
--- a/tests/asm-c-connect-1.c
+++ b/tests/asm-c-connect-1.c
@@ -19,7 +19,12 @@ static int __USED x1_c (void)
     return 1;
 }
 
+#if __i386__
 asm(".text;"_"x1: call "_"x1_c; ret");
+#else
+/* Keep stack aligned */
+asm(".text;"_"x1: sub $8,%rsp; call "_"x1_c; add $8,%rsp; ret");
+#endif
 
 void callx4(void);
 void callx5_again(void);
diff --git a/tests/asm-c-connect-2.c b/tests/asm-c-connect-2.c
index e2b4b595..654db0e3 100644
--- a/tests/asm-c-connect-2.c
+++ b/tests/asm-c-connect-2.c
@@ -22,7 +22,12 @@ int x3(void)
 /* That callx4 is defined globally (as if ".globl callx4")
    is a TCC extension.  GCC doesn't behave like this.  */
 void callx4(void);
+#if __i386__
 __asm__(_"callx4: call "_"x4; ret;"
+#else
+/* Keep stack aligned */
+__asm__(_"callx4: sub $8,%rsp; call "_"x4; add $8,%rsp; ret;"
+#endif
 #ifndef __TINYC__
     " .global "_"callx4"
 #endif
diff --git a/tests/pp/Makefile b/tests/pp/Makefile
index 4785db3b..224c866d 100644
--- a/tests/pp/Makefile
+++ b/tests/pp/Makefile
@@ -12,7 +12,7 @@ TESTS = $(call files,c) $(call files,S)
 
 all test testspp.all: $(sort $(TESTS))
 
-DIFF_OPTS = -Nu -b -B
+DIFF_OPTS = -Nu -b
 
 # Filter source directory in warnings/errors (out-of-tree builds)
 FILTER = 2>&1 | sed 's,$(SRC)/,,g'