diff --git a/tccgen.c b/tccgen.c
index a6181b0..4bac2b9 100644
--- a/tccgen.c
+++ b/tccgen.c
@@ -7413,7 +7413,7 @@ static void decl_initializer_alloc(CType *type, AttributeDef *ad, int r,
     if ((r & VT_VALMASK) == VT_LOCAL) {
         sec = NULL;
 #ifdef CONFIG_TCC_BCHECK
-        if (bcheck && (type->t & VT_ARRAY)) {
+        if (bcheck) {
             loc--;
         }
 #endif
@@ -7422,8 +7422,13 @@ static void decl_initializer_alloc(CType *type, AttributeDef *ad, int r,
 #ifdef CONFIG_TCC_BCHECK
         /* handles bounds */
         /* XXX: currently, since we do only one pass, we cannot track
-           '&' operators, so we add only arrays */
-        if (bcheck && (type->t & VT_ARRAY)) {
+           '&' operators.
+	   So add all values because address could be taken or
+	   a struct/union could contain an array.
+	   This makes code very slow!
+	   In the future we might remove the values from the bounds_section
+	   if addres is not taken. */
+        if (bcheck) {
             addr_t *bounds_ptr;
             /* add padding between regions */
             loc--;
diff --git a/x86_64-gen.c b/x86_64-gen.c
index cc66b60..38a7787 100644
--- a/x86_64-gen.c
+++ b/x86_64-gen.c
@@ -654,6 +654,8 @@ ST_FUNC void gen_bounded_ptr_add(void)
     /* save all temporary registers */
     save_regs(0);
 
+    o(0x51525657); /* push $rdi/%rsi/%rdx/%rcx */
+
     /* prepare fast x86_64 function call */
     gv(RC_RAX);
     o(0xc68948); // mov  %rax,%rsi ## second arg in %rsi, this must be size
@@ -670,6 +672,7 @@ ST_FUNC void gen_bounded_ptr_add(void)
     vtop++;
     vtop->r = TREG_RAX | VT_BOUNDED;
 
+    o(0x5f5e5a59); /* pop $rcx/$rdx/$rsi/%rdi */
 
     /* relocation offset of the bounding function call point */
     vtop->c.i = (cur_text_section->reloc->data_offset - sizeof(ElfW(Rela)));
@@ -1940,6 +1943,7 @@ void gen_opf(int op)
                 v1.c.i = fc;
                 load(r, &v1);
                 fc = 0;
+	        vtop->r = r = r | VT_LVAL;
             }
 
             if (op == TOK_EQ || op == TOK_NE) {
@@ -2007,6 +2011,7 @@ void gen_opf(int op)
                 v1.c.i = fc;
                 load(r, &v1);
                 fc = 0;
+	        vtop->r = r = r | VT_LVAL;
             }
             
             assert(!(vtop[-1].r & VT_LVAL));