diff --git a/Makefile b/Makefile
index 9cf6f7e0..7679b688 100644
--- a/Makefile
+++ b/Makefile
@@ -16,7 +16,6 @@ endif
 
 LIBTCC = libtcc.a
 LIBTCC1 = libtcc1.a
-LIBTCCB1 = libtccb1.a
 LINK_LIBTCC =
 LIBS =
 CFLAGS += -I$(TOP)
@@ -87,7 +86,7 @@ ifeq ($(INCLUDED),no)
 # running top Makefile
 
 PROGS = tcc$(EXESUF)
-TCCLIBS = $(LIBTCC1) $(LIBTCCB1) $(LIBTCC) $(LIBTCCDEF)
+TCCLIBS = $(LIBTCC1) $(LIBTCC) $(LIBTCCDEF)
 TCCDOCS = tcc.1 tcc-doc.html tcc-doc.info
 
 all: $(PROGS) $(TCCLIBS) $(TCCDOCS)
@@ -101,15 +100,11 @@ TCC_X += riscv64
 LIBTCC1_X = i386 x86_64 i386-win32 x86_64-win32 x86_64-osx arm arm64 arm-wince
 LIBTCC1_X += riscv64
 
-# cross libtccb1.a targets to build
-LIBTCCB1_X = i386 x86_64 i386-win32 x86_64-win32 
-
 PROGS_CROSS = $(foreach X,$(TCC_X),$X-tcc$(EXESUF))
 LIBTCC1_CROSS = $(foreach X,$(LIBTCC1_X),$X-libtcc1.a)
-LIBTCCB1_CROSS = $(foreach X,$(LIBTCCB1_X),$X-libtcc1.a)
 
 # build cross compilers & libs
-cross: $(LIBTCC1_CROSS) $(LIBTCCB1_CROSS) $(PROGS_CROSS)
+cross: $(LIBTCC1_CROSS) $(PROGS_CROSS)
 
 # build specific cross compiler & lib
 cross-%: %-tcc$(EXESUF) %-libtcc1.a ;
@@ -153,10 +148,10 @@ DEFINES += $(DEF-$(or $(findstring win,$T),unx))
 
 ifneq ($(X),)
 ifeq ($(CONFIG_WIN32),yes)
-DEF-win += -DTCC_LIBTCC1="\"$(X)libtcc1.a\"" -DTCC_LIBTCCB1="\"$(X)libtccb1.a\""
-DEF-unx += -DTCC_LIBTCC1="\"lib/$(X)libtcc1.a\"" -DTCC_LIBTCCB1="\"lib/$(X)libtccb1.a\""
+DEF-win += -DTCC_LIBTCC1="\"$(X)libtcc1.a\""
+DEF-unx += -DTCC_LIBTCC1="\"lib/$(X)libtcc1.a\""
 else
-DEF-all += -DTCC_LIBTCC1="\"$(X)libtcc1.a\"" -DTCC_LIBTCCB1="\"$(X)libtccb1.a\""
+DEF-all += -DTCC_LIBTCC1="\"$(X)libtcc1.a\""
 DEF-win += -DCONFIG_TCCDIR="\"$(tccdir)/win32\""
 endif
 endif
@@ -281,8 +276,6 @@ STRIP_yes = -s
 
 LIBTCC1_W = $(filter %-win32-libtcc1.a %-wince-libtcc1.a,$(LIBTCC1_CROSS))
 LIBTCC1_U = $(filter-out $(LIBTCC1_W),$(LIBTCC1_CROSS))
-LIBTCCB1_W = $(filter %-win32-libtccb1.a %-wince-libtccb1.a,$(LIBTCCB1_CROSS))
-LIBTCCB1_U = $(filter-out $(LIBTCCB1_W),$(LIBTCCB1_CROSS))
 IB = $(if $1,mkdir -p $2 && $(INSTALLBIN) $1 $2)
 IBw = $(call IB,$(wildcard $1),$2)
 IF = $(if $1,mkdir -p $2 && $(INSTALL) $1 $2)
@@ -292,8 +285,7 @@ IR = mkdir -p $2 && cp -r $1/. $2
 # install progs & libs
 install-unx:
 	$(call IBw,$(PROGS) $(PROGS_CROSS),"$(bindir)")
-	$(call IFw,$(LIBTCC1) $(LIBTCC1_U),"$(tccdir)")
-	$(call IFw,$(LIBTCCB1) $(LIBTCCB1_U),"$(tccdir)")
+	$(call IFw,$(LIBTCC1) bcheck.o $(LIBTCC1_U),"$(tccdir)")
 	$(call IF,$(TOPSRC)/include/*.h $(TOPSRC)/tcclib.h,"$(tccdir)/include")
 	$(call $(if $(findstring .so,$(LIBTCC)),IBw,IFw),$(LIBTCC),"$(libdir)")
 	$(call IF,$(TOPSRC)/libtcc.h,"$(includedir)")
@@ -318,8 +310,7 @@ uninstall-unx:
 install-win:
 	$(call IBw,$(PROGS) $(PROGS_CROSS) $(subst libtcc.a,,$(LIBTCC)),"$(bindir)")
 	$(call IF,$(TOPSRC)/win32/lib/*.def,"$(tccdir)/lib")
-	$(call IFw,libtcc1.a $(LIBTCC1_W),"$(tccdir)/lib")
-	$(call IFw,libtccb1.a $(LIBTCCB1_W),"$(tccdir)/lib")
+	$(call IFw,libtcc1.a bcheck.o $(LIBTCC1_W),"$(tccdir)/lib")
 	$(call IF,$(TOPSRC)/include/*.h $(TOPSRC)/tcclib.h,"$(tccdir)/include")
 	$(call IR,$(TOPSRC)/win32/include,"$(tccdir)/include")
 	$(call IR,$(TOPSRC)/win32/examples,"$(tccdir)/examples")
diff --git a/i386-gen.c b/i386-gen.c
index a3ac548d..6ff096ad 100644
--- a/i386-gen.c
+++ b/i386-gen.c
@@ -425,6 +425,11 @@ ST_FUNC void gfunc_call(int nb_args)
     int size, align, r, args_size, i, func_call;
     Sym *func_sym;
     
+#ifdef CONFIG_TCC_BCHECK
+    if (tcc_state->do_bounds_check)
+        gbound_args(nb_args);
+#endif
+
     args_size = 0;
     for(i = 0;i < nb_args; i++) {
         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
@@ -1051,119 +1056,18 @@ ST_FUNC void ggoto(void)
 
 /* bound check support functions */
 #ifdef CONFIG_TCC_BCHECK
-
-ST_FUNC void tcc_add_bcheck(TCCState *s1, Section *bound_sec, Section *sym_sec)
-{
-    addr_t *ptr;
-    int loc_glob;
-    int sym_index;
-    int bsym_index;
-
-    if (0 == s1->do_bounds_check)
-        return;
-    /* XXX: add an object file to do that */
-    ptr = section_ptr_add(bound_sec, sizeof(*ptr));
-    *ptr = 0;
-    loc_glob = s1->output_type != TCC_OUTPUT_MEMORY ? STB_LOCAL : STB_GLOBAL;
-    bsym_index = set_elf_sym(sym_sec, 0, 0,
-                             ELFW(ST_INFO)(loc_glob, STT_NOTYPE), 0,
-                             bound_sec->sh_num, "__bounds_start");
-    /* pull bcheck.o from libtcc1.a */
-    sym_index = set_elf_sym(sym_sec, 0, 0,
-                            ELFW(ST_INFO)(STB_GLOBAL, STT_NOTYPE), 0,
-                            SHN_UNDEF, "__bound_init");
-    if (s1->output_type != TCC_OUTPUT_MEMORY) {
-        /* add 'call __bound_init()' in .init section */
-        Section *init_section = find_section(s1, ".init");
-        unsigned char *pinit;
-#ifdef TCC_TARGET_PE
-        pinit = section_ptr_add(init_section, 3);
-        pinit[0] = 0x55;        /* push %rbp */
-        pinit[1] = 0x89;        /* mov %esp,%ebp */
-        pinit[2] = 0xe5;
-#endif
-        pinit = section_ptr_add(init_section, 5);
-        pinit[0] = 0xe8;
-        write32le(pinit + 1, -4);
-        put_elf_reloc(sym_sec, init_section,
-            init_section->data_offset - 4, R_386_PC32, sym_index);
-            /* R_386_PC32 = R_X86_64_PC32 = 2 */
-        pinit = section_ptr_add(init_section, 6);
-        pinit[0] = 0xb8;        /* mov xx,%eax */
-        write32le(pinit + 1, 0);
-        pinit[5] = 0x50;        /* push %eax */
-        put_elf_reloc(sym_sec, init_section,
-                init_section->data_offset - 5, R_386_32, bsym_index);
-        sym_index = set_elf_sym(sym_sec, 0, 0,
-                        ELFW(ST_INFO)(STB_GLOBAL, STT_NOTYPE), 0,
-                        SHN_UNDEF, "__bounds_add_static_var");
-        pinit = section_ptr_add(init_section, 5);
-        pinit[0] = 0xe8;
-        write32le(pinit + 1, -4);
-        put_elf_reloc(sym_sec, init_section,
-            init_section->data_offset - 4, R_386_PC32, sym_index);
-                /* R_386_PC32 = R_X86_64_PC32 = 2 */
-        pinit = section_ptr_add(init_section, 3);
-        pinit[0] = 0x83;        /* add  $0x4,%esp */
-        pinit[1] = 0xc4;
-        pinit[2] = 0x04;
-#ifdef TCC_TARGET_PE
-        {
-            int init_index = set_elf_sym(sym_sec,
-                                         0, 0,
-                                         ELFW(ST_INFO)(STB_GLOBAL, STT_NOTYPE), 0,
-                                         init_section->sh_num, "__init_start");
-            Sym sym;
-            init_section->sh_flags |= SHF_EXECINSTR;
-            pinit = section_ptr_add(init_section, 2);
-            pinit[0] = 0xc9;        /* leave */
-            pinit[1] = 0xc3;        /* ret */
-            sym.c = init_index;
-            add_init_array (s1, &sym);
-        }
-#endif
-    }
-}
-
 /* generate a bounded pointer addition */
 ST_FUNC void gen_bounded_ptr_add(void)
 {
-    /* save all temporary registers */
-    save_regs(0);
-    /* prepare fast i386 function call (args in eax and edx) */
-    gv2(RC_EAX, RC_EDX);
-    vtop -= 2;
-    /* add line, filename */
-    {
-        static addr_t offset;
-        static char last_filename[1024];
-        Sym *sym_data;
-
-        if (strcmp (last_filename, file->filename) != 0) {
-            void *ptr;
-            int len = strlen (file->filename) + 1;
-
-            offset = data_section->data_offset;
-            ptr = section_ptr_add(data_section, len);
-            memcpy (ptr, file->filename, len);
-            memcpy (last_filename, file->filename, len);
-        }
-        o(0xb9);   /* mov $xx,%ecx */
-        gen_le32 (0);
-        sym_data = get_sym_ref(&char_pointer_type, data_section,
-                               offset, data_section->data_offset);
-        greloca(cur_text_section, sym_data, ind - 4, R_386_32, 0);
-        o(0x51);   /* push %ecx */
-    }
-    o(0xb9);       /* mov $xx,%ecx */
-    gen_le32 (file->line_num);
-    /* do a fast function call */
-    gen_static_call(TOK___bound_ptr_add);
-    o(0x04c483);   /* add $4,%esp */
+    vpush_global_sym(&func_old_type, TOK___bound_ptr_add);
+    vrott(3);
+    gfunc_call(2);
+    vpushi(0);
     /* returned pointer is in eax */
-    vtop++;
     vtop->r = TREG_EAX | VT_BOUNDED;
-    /* address of bounding function call point */
+    if (nocode_wanted)
+        return;
+    /* relocation offset of the bounding function call point */
     vtop->c.i = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
 }
 
@@ -1176,6 +1080,9 @@ ST_FUNC void gen_bounded_ptr_deref(void)
     Elf32_Rel *rel;
     Sym *sym;
 
+    if (nocode_wanted)
+        return;
+
     size = 0;
     /* XXX: put that code in generic part of tcc */
     if (!is_float(vtop->type.t)) {
@@ -1194,17 +1101,18 @@ ST_FUNC void gen_bounded_ptr_deref(void)
     case 12: func = TOK___bound_ptr_indir12; break;
     case 16: func = TOK___bound_ptr_indir16; break;
     default:
-        tcc_error("unhandled size when dereferencing bounded pointer");
-        func = 0;
-        break;
+        /* may happen with struct member access */
+        return;
+        //tcc_error("unhandled size when dereferencing bounded pointer");
+        //func = 0;
+        //break;
     }
-
-    /* patch relocation */
-    /* XXX: find a better solution ? */
-    rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.i);
     sym = external_global_sym(func, &func_old_type);
     if (!sym->c)
         put_extern_sym(sym, NULL, 0, 0);
+    /* patch relocation */
+    /* XXX: find a better solution ? */
+    rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.i);
     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
 }
 #endif
diff --git a/i386-link.c b/i386-link.c
index e4929b43..23653b29 100644
--- a/i386-link.c
+++ b/i386-link.c
@@ -235,4 +235,79 @@ void relocate(TCCState *s1, ElfW_Rel *rel, int type, unsigned char *ptr, addr_t
     }
 }
 
+#ifdef CONFIG_TCC_BCHECK
+ST_FUNC void tcc_add_bcheck(TCCState *s1)
+{
+    addr_t *ptr;
+    int loc_glob;
+    int sym_index;
+    int bsym_index;
+
+    if (0 == s1->do_bounds_check)
+        return;
+    /* XXX: add an object file to do that */
+    ptr = section_ptr_add(bounds_section, sizeof(*ptr));
+    *ptr = 0;
+    loc_glob = s1->output_type != TCC_OUTPUT_MEMORY ? STB_LOCAL : STB_GLOBAL;
+    bsym_index = set_elf_sym(symtab_section, 0, 0,
+                             ELFW(ST_INFO)(loc_glob, STT_NOTYPE), 0,
+                             bounds_section->sh_num, "__bounds_start");
+    /* pull bcheck.o from libtcc1.a */
+    sym_index = set_elf_sym(symtab_section, 0, 0,
+                            ELFW(ST_INFO)(STB_GLOBAL, STT_NOTYPE), 0,
+                            SHN_UNDEF, "__bound_init");
+    if (s1->output_type != TCC_OUTPUT_MEMORY) {
+        /* add 'call __bound_init()' in .init section */
+        Section *init_section = find_section(s1, ".init");
+        unsigned char *pinit;
+#ifdef TCC_TARGET_PE
+        pinit = section_ptr_add(init_section, 3);
+        pinit[0] = 0x55;        /* push %rbp */
+        pinit[1] = 0x89;        /* mov %esp,%ebp */
+        pinit[2] = 0xe5;
+#endif
+        pinit = section_ptr_add(init_section, 5);
+        pinit[0] = 0xe8;
+        write32le(pinit + 1, -4);
+        put_elf_reloc(symtab_section, init_section,
+            init_section->data_offset - 4, R_386_PC32, sym_index);
+            /* R_386_PC32 = R_X86_64_PC32 = 2 */
+        pinit = section_ptr_add(init_section, 6);
+        pinit[0] = 0xb8;        /* mov xx,%eax */
+        write32le(pinit + 1, 0);
+        pinit[5] = 0x50;        /* push %eax */
+        put_elf_reloc(symtab_section, init_section,
+                init_section->data_offset - 5, R_386_32, bsym_index);
+        sym_index = set_elf_sym(symtab_section, 0, 0,
+                        ELFW(ST_INFO)(STB_GLOBAL, STT_NOTYPE), 0,
+                        SHN_UNDEF, "__bounds_add_static_var");
+        pinit = section_ptr_add(init_section, 5);
+        pinit[0] = 0xe8;
+        write32le(pinit + 1, -4);
+        put_elf_reloc(symtab_section, init_section,
+            init_section->data_offset - 4, R_386_PC32, sym_index);
+                /* R_386_PC32 = R_X86_64_PC32 = 2 */
+        pinit = section_ptr_add(init_section, 3);
+        pinit[0] = 0x83;        /* add  $0x4,%esp */
+        pinit[1] = 0xc4;
+        pinit[2] = 0x04;
+#ifdef TCC_TARGET_PE
+        {
+            int init_index = set_elf_sym(symtab_section,
+                                         0, 0,
+                                         ELFW(ST_INFO)(STB_GLOBAL, STT_NOTYPE), 0,
+                                         init_section->sh_num, "__init_start");
+            Sym sym;
+            init_section->sh_flags |= SHF_EXECINSTR;
+            pinit = section_ptr_add(init_section, 2);
+            pinit[0] = 0xc9;        /* leave */
+            pinit[1] = 0xc3;        /* ret */
+            sym.c = init_index;
+            add_init_array (s1, &sym);
+        }
+#endif
+    }
+}
+#endif
+
 #endif /* !TARGET_DEFS_ONLY */
diff --git a/lib/Makefile b/lib/Makefile
index 2a9d7abe..b8c878bb 100644
--- a/lib/Makefile
+++ b/lib/Makefile
@@ -7,8 +7,6 @@ include $(TOP)/Makefile
 VPATH = $(TOPSRC)/lib $(TOPSRC)/win32/lib
 T = $(or $(CROSS_TARGET),$(NATIVE_TARGET),unknown)
 X = $(if $(CROSS_TARGET),$(CROSS_TARGET)-)
-BIN = $(TOP)/$(X)libtcc1.a
-BINB = $(TOP)/$(X)libtccb1.a
 
 XTCC ?= $(TOP)/$(X)tcc$(EXESUF)
 XCC = $(XTCC)
@@ -20,8 +18,6 @@ XCFG = $(or $(findstring -win,$T),-unx)
 
 # in order to use gcc, tyoe: make <target>-libtcc1-usegcc=yes
 arm-libtcc1-usegcc ?= no
-x86_64-libtcc1-usegcc ?= no
-i386-libtcc1-usegcc ?= no
 
 ifeq "$($(T)-libtcc1-usegcc)" "yes"
  XCC = $(CC)
@@ -42,8 +38,6 @@ ifdef CONFIG_OSX
  XFLAGS += -D_ANSI_SOURCE
 endif
 
-XFLAGS += -g
-
 I386_O = libtcc1.o alloca86.o alloca86-bt.o
 X86_64_O = libtcc1.o alloca86_64.o alloca86_64-bt.o
 ARM_O = libtcc1.o armeabi.o alloca-arm.o armflush.o
@@ -51,11 +45,11 @@ ARM64_O = lib-arm64.o
 RISCV64_O = lib-arm64.o
 WIN_O = crt1.o crt1w.o wincrt1.o wincrt1w.o dllcrt1.o dllmain.o
 
-OBJ-i386 = $(I386_O) $(DSO_O)
-OBJ-x86_64 = $(X86_64_O) va_list.o $(DSO_O)
+OBJ-i386 = $(I386_O) $(BCHECK_O) $(DSO_O)
+OBJ-x86_64 = $(X86_64_O) va_list.o $(BCHECK_O) $(DSO_O)
 OBJ-x86_64-osx = $(X86_64_O) va_list.o
-OBJ-i386-win32 = $(I386_O) chkstk.o $(WIN_O)
-OBJ-x86_64-win32 = $(X86_64_O) chkstk.o (WIN_O)
+OBJ-i386-win32 = $(I386_O) chkstk.o bcheck.o $(WIN_O)
+OBJ-x86_64-win32 = $(X86_64_O) chkstk.o bcheck.o $(WIN_O)
 OBJ-arm64 = $(ARM64_O) $(DSO_O)
 OBJ-arm = $(ARM_O) $(DSO_O)
 OBJ-arm-fpa = $(ARM_O) $(DSO_O)
@@ -66,27 +60,14 @@ OBJ-arm-eabihf = $(ARM_O) $(DSO_O)
 OBJ-arm-wince = $(ARM_O) $(WIN_O)
 OBJ-riscv64 = $(RISCV64_O) $(DSO_O)
 
-OBJB-i386 = $(BCHECK_O)
-OBJB-x86_64 = $(BCHECK_O)
-OBJB-x86_64-osx = dummy.o
-OBJB-i386-win32 = bcheck.o
-OBJB-x86_64-win32 = bcheck.o
-OBJB-arm64 = dummy.o
-OBJB-arm = dummy.o
-OBJB-arm-fpa = dummy.o
-OBJB-arm-fpa-ld = dummy.o
-OBJB-arm-vfp = dummy.o
-OBJB-arm-eabi = dummy.o
-OBJB-arm-eabihf = dummy.o
-OBJB-arm-wince = dummy.o
-OBJB-riscv64 = dummy.o
+OBJ-extra = $(filter bcheck.o,$(OBJ-$T))
+OBJ-libtcc1 = $(addprefix $(X),$(filter-out $(OBJ-extra),$(OBJ-$T)))
 
-all: $(BIN) $(BINB)
+ALL = $(addprefix $(TOP)/,$(X)libtcc1.a $(OBJ-extra))
 
-$(BIN) : $(patsubst %.o,$(X)%.o,$(OBJ-$T))
-	$(XAR) rcs $@ $^
+all: $(ALL)
 
-$(BINB) : $(patsubst %.o,$(X)%.o,$(OBJB-$T))
+$(TOP)/$(X)libtcc1.a : $(OBJ-libtcc1)
 	$(XAR) rcs $@ $^
 
 $(X)%.o : %.c
@@ -95,8 +76,12 @@ $(X)%.o : %.c
 $(X)%.o : %.S
 	$(XCC) -c $< -o $@ $(XFLAGS)
 
+$(TOP)/%.o : %.c
+	$(XCC) -c $< -o $@ $(XFLAGS)
+
+$(TOP)/bcheck.o : XFLAGS += -g
+
 $(X)crt1w.o : crt1.c
 $(X)wincrt1w.o : wincrt1.c
-
 clean :
-	rm -f *.a *.o $(BIN) $(BINB)
+	rm -f *.a *.o $(ALL)
diff --git a/lib/alloca86_64-bt.S b/lib/alloca86_64-bt.S
index 15bca090..e1bac5f4 100644
--- a/lib/alloca86_64-bt.S
+++ b/lib/alloca86_64-bt.S
@@ -5,44 +5,25 @@
 __bound_alloca:
 
 #ifdef _WIN32
-    pop     %rdx
-    mov     %rcx,%rax
-    add     $15,%rax
-    and     $-16,%rax
-    jz      p3
-
-p1:
-    cmp     $4096,%rax
-    jbe     p2
-    test    %rax,-4096(%rsp)
-    sub     $4096,%rsp
-    sub     $4096,%rax
-    jmp p1
-p2:
-
-    sub     %rax,%rsp
-    mov     %rsp,%rax
-
-    push    %rdx
+    inc %rcx            # add one extra to separate regions
+    jmp alloca
+.globl __bound_alloca_nr
+__bound_alloca_nr:
+    dec     %rcx
     push    %rax
     mov     %rcx,%rdx
     mov     %rax,%rcx
-    sub     $20,%rsp
-    call   __bound_new_region
-    add     $20,%rsp
+    sub     $32,%rsp
+    call    __bound_new_region
+    add     $32,%rsp
     pop     %rax
-    pop     %rdx
-
-    add     $32,%rax
-p3:
-    push    %rdx
     ret
 #else
     pop     %rdx
     mov     %rdi,%rax
     mov     %rax,%rsi	# size, a second parm to the __bound_new_region
 
-    add     $15,%rax
+    add     $15 + 1,%rax  # add one extra to separate regions
     and     $-16,%rax
     jz      p3
 
diff --git a/lib/alloca86_64.S b/lib/alloca86_64.S
index 8f89fe48..a4aa173c 100644
--- a/lib/alloca86_64.S
+++ b/lib/alloca86_64.S
@@ -24,12 +24,8 @@ p1:
     jmp p1
 p2:
 #endif
-
     sub     %rax,%rsp
     mov     %rsp,%rax
-#ifdef _WIN32
-    add     $32,%rax
-#endif
 p3:
     push    %rdx
     ret
diff --git a/lib/bcheck.c b/lib/bcheck.c
index 9c913784..ab8f4675 100644
--- a/lib/bcheck.c
+++ b/lib/bcheck.c
@@ -205,6 +205,9 @@ void __bound_checking (int no_check)
     no_checking = no_check;
 }
 
+#define no_FASTCALL
+//#define no_checking 1
+
 /* print a bound error message */
 static void bound_error(const char *fmt, ...)
 {
@@ -221,8 +224,7 @@ static void bound_alloc_error(void)
 
 /* return '(p + offset)' for pointer arithmetic (a pointer can reach
    the end of a region in this case */
-void * FASTCALL __bound_ptr_add(void *p, size_t offset,
-                                size_t line, const char *filename)
+void * no_FASTCALL __bound_ptr_add(void *p, size_t offset)
 {
     size_t addr = (size_t)p;
 
@@ -230,8 +232,8 @@ void * FASTCALL __bound_ptr_add(void *p, size_t offset,
         return p + offset;
     }
 
-    dprintf(stderr, "%s %s (%s:%u): %p 0x%x\n",
-            __FILE__, __FUNCTION__, filename, line, p, (unsigned)offset);
+    dprintf(stderr, "%s %s : %p 0x%x\n",
+            __FILE__, __FUNCTION__, p, (unsigned)offset);
 
     WAIT_SEM ();
     INCR_COUNT(bound_ptr_add_count);
@@ -250,8 +252,10 @@ void * FASTCALL __bound_ptr_add(void *p, size_t offset,
         if (addr <= tree->size) {
             addr += offset;
             if (tree->is_invalid || addr > tree->size) {
-                fprintf(stderr,"%s %s (%s:%u): %p is outside of the region\n",
-                        __FILE__, __FUNCTION__, filename, line, p + offset);
+        #if 0
+                fprintf(stderr,"%s %s : %p is outside of the region\n",
+                        __FILE__, __FUNCTION__, p + offset);
+        #endif
                 if (never_fatal == 0) {
                     POST_SEM ();
                     return INVALID_POINTER; /* return an invalid pointer */
@@ -266,16 +270,15 @@ void * FASTCALL __bound_ptr_add(void *p, size_t offset,
 /* return '(p + offset)' for pointer indirection (the resulting must
    be strictly inside the region */
 #define BOUND_PTR_INDIR(dsize)                                                 \
-void * FASTCALL __bound_ptr_indir ## dsize (void *p, size_t offset,            \
-                                            size_t line, const char *filename) \
+void * no_FASTCALL __bound_ptr_indir ## dsize (void *p, size_t offset)         \
 {                                                                              \
     size_t addr = (size_t)p;                                                   \
                                                                                \
     if (no_checking) {                                                         \
         return p + offset;                                                     \
     }                                                                          \
-    dprintf(stderr, "%s %s (%s:%u): %p 0x%x start\n",                          \
-            __FILE__, __FUNCTION__, filename, line,  p, (unsigned)offset);     \
+    dprintf(stderr, "%s %s : %p 0x%x start\n",                                 \
+            __FILE__, __FUNCTION__, p, (unsigned)offset);                      \
     WAIT_SEM ();                                                               \
     INCR_COUNT(bound_ptr_indir ## dsize ## _count);                            \
     if (tree) {                                                                \
@@ -293,8 +296,8 @@ void * FASTCALL __bound_ptr_indir ## dsize (void *p, size_t offset,            \
         if (addr <= tree->size) {                                              \
             addr += offset + dsize;                                            \
             if (tree->is_invalid || addr > tree->size) {                       \
-                fprintf(stderr,"%s %s (%s:%u): %p is outside of the region\n", \
-                    __FILE__, __FUNCTION__, filename, line, p + offset);       \
+                fprintf(stderr,"%s %s : %p is outside of the region\n",        \
+                    __FILE__, __FUNCTION__, p + offset);                       \
                 if (never_fatal == 0) {                                        \
                     POST_SEM ();                                               \
                     return INVALID_POINTER; /* return an invalid pointer */    \
@@ -907,7 +910,7 @@ static void __bound_check(const void *p, size_t size, const char *function)
         return;
     if (size == 0)
         return;
-    p = __bound_ptr_add((void *)p, size, 0, function);
+    p = __bound_ptr_add((void *)p, size);
     if (p == INVALID_POINTER)
         bound_error("invalid pointer");
 }
@@ -959,7 +962,7 @@ int __bound_strlen(const char *s)
     INCR_COUNT(bound_strlen_count);
     while (*p++);
     len = (p - s) - 1;
-    p = __bound_ptr_indir1((char *)s, len, 0, "strlen");
+    p = __bound_ptr_indir1((char *)s, len);
     if (p == INVALID_POINTER)
         bound_error("bad pointer in strlen()");
     return len;
diff --git a/lib/dummy.c b/lib/dummy.c
deleted file mode 100644
index c1548940..00000000
--- a/lib/dummy.c
+++ /dev/null
@@ -1 +0,0 @@
-static char dummy;
diff --git a/libtcc.c b/libtcc.c
index 0d58a30b..3094d2df 100644
--- a/libtcc.c
+++ b/libtcc.c
@@ -1545,7 +1545,6 @@ static const TCCOption tcc_options[] = {
 #endif
 #ifdef CONFIG_TCC_BCHECK
     { "b", TCC_OPTION_b, 0 },
-    { "ba", TCC_OPTION_ba, 0 },
 #endif
     { "g", TCC_OPTION_g, TCC_OPTION_HAS_ARG | TCC_OPTION_NOSEP },
     { "c", TCC_OPTION_c, 0 },
@@ -1802,9 +1801,6 @@ reparse:
             s->do_bounds_check = 1;
             s->do_debug = 1;
             break;
-        case TCC_OPTION_ba:
-            s->do_bounds_check_address = 1;
-            break;
 #endif
         case TCC_OPTION_g:
             s->do_debug = 1;
diff --git a/tcc.c b/tcc.c
index ab054bfc..63250c69 100644
--- a/tcc.c
+++ b/tcc.c
@@ -59,7 +59,6 @@ static const char help[] =
     "  -g          generate runtime debug info\n"
 #ifdef CONFIG_TCC_BCHECK
     "  -b          compile with built-in memory and bounds checker (implies -g)\n"
-    "  -ba         Enable better address checking with bounds checker\n"
 #endif
 #ifdef CONFIG_TCC_BACKTRACE
     "  -bt N       show N callers in stack traces\n"
diff --git a/tcc.h b/tcc.h
index 127a19a9..13579894 100644
--- a/tcc.h
+++ b/tcc.h
@@ -292,9 +292,6 @@ extern long double strtold (const char *__nptr, char **__endptr);
 #ifndef TCC_LIBTCC1
 # define TCC_LIBTCC1 "libtcc1.a"
 #endif
-#ifndef TCC_LIBTCCB1
-# define TCC_LIBTCCB1 "libtccb1.a"
-#endif
 
 /* library to use with CONFIG_USE_LIBGCC instead of libtcc1.a */
 #if defined CONFIG_USE_LIBGCC && !defined TCC_LIBGCC
@@ -720,7 +717,6 @@ struct TCCState {
 #ifdef CONFIG_TCC_BCHECK
     /* compile with built-in memory and bounds checker */
     unsigned char do_bounds_check;
-    unsigned char do_bounds_check_address;
 #endif
 #ifdef TCC_TARGET_ARM
     enum float_abi float_abi; /* float ABI of the generated code*/
@@ -1250,7 +1246,7 @@ ST_FUNC int tcc_add_crt(TCCState *s, const char *filename);
 #endif
 ST_FUNC int tcc_add_dll(TCCState *s, const char *filename, int flags);
 #ifdef CONFIG_TCC_BCHECK
-ST_FUNC void tcc_add_bcheck(TCCState *s1, Section *bound_sec, Section *sym_sec);
+ST_FUNC void tcc_add_bcheck(TCCState *s1);
 #endif
 ST_FUNC void tcc_add_pragma_libs(TCCState *s1);
 PUB_FUNC int tcc_add_library_err(TCCState *s, const char *f);
@@ -1437,6 +1433,9 @@ ST_FUNC Sym *get_sym_ref(CType *type, Section *sec, unsigned long offset, unsign
 #if defined TCC_TARGET_X86_64 && !defined TCC_TARGET_PE
 ST_FUNC int classify_x86_64_va_arg(CType *ty);
 #endif
+#ifdef CONFIG_TCC_BCHECK
+ST_FUNC void gbound_args(int nb_args);
+#endif
 
 /* ------------ tccelf.c ------------ */
 
diff --git a/tccelf.c b/tccelf.c
index 20fc20c5..f5f55b1c 100644
--- a/tccelf.c
+++ b/tccelf.c
@@ -1321,7 +1321,7 @@ ST_FUNC void tcc_add_runtime(TCCState *s1)
 {
     s1->filetype = 0;
 #ifdef CONFIG_TCC_BCHECK
-    tcc_add_bcheck(s1, bounds_section, symtab_section);
+    tcc_add_bcheck(s1);
 #endif
     tcc_add_pragma_libs(s1);
 #ifndef TCC_TARGET_PE
@@ -1340,7 +1340,7 @@ ST_FUNC void tcc_add_runtime(TCCState *s1)
         if (s1->do_bounds_check && s1->output_type != TCC_OUTPUT_DLL) {
             tcc_add_library_err(s1, "pthread");
             tcc_add_library_err(s1, "dl");
-            tcc_add_support(s1, TCC_LIBTCCB1);
+            tcc_add_support(s1, "bcheck.o");
         }
 #endif
         tcc_add_support(s1, TCC_LIBTCC1);
diff --git a/tccgen.c b/tccgen.c
index 32357f55..01473fff 100644
--- a/tccgen.c
+++ b/tccgen.c
@@ -1230,6 +1230,11 @@ ST_FUNC void save_reg_upstack(int r, int n)
                     type = &int_type;
 #endif
                 size = type_size(type, &align);
+#ifdef CONFIG_TCC_BCHECK
+                if (tcc_state->do_bounds_check)
+                    l = loc = (loc - size) & -align;
+                else
+#endif
                 l=get_temp_local_var(size,align);
                 sv.type.t = type->t;
                 sv.r = VT_LOCAL | VT_LVAL;
@@ -1417,7 +1422,7 @@ static void gbound(void)
 
     vtop->r &= ~VT_MUSTBOUND;
     /* if lvalue, then use checking code before dereferencing */
-    if ((vtop->r & VT_LVAL) && !nocode_wanted) {
+    if (vtop->r & VT_LVAL) {
         /* if not VT_BOUNDED value, then make one */
         if (!(vtop->r & VT_BOUNDED)) {
             lval_type = vtop->r & (VT_LVAL_TYPE | VT_LVAL);
@@ -1434,6 +1439,19 @@ static void gbound(void)
         gen_bounded_ptr_deref();
     }
 }
+
+/* we need to call __bound_ptr_add before we start to load function
+   args into registers */
+ST_FUNC void gbound_args(int nb_args)
+{
+    int i;
+    for (i = 1; i <= nb_args; ++i)
+        if (vtop[1 - i].r & VT_MUSTBOUND) {
+            vrotb(i);
+            gbound();
+            vrott(i);
+        }
+}
 #endif
 
 static void incr_bf_adr(int o)
@@ -2499,25 +2517,7 @@ redo:
             }
             gen_op('*');
 #ifdef CONFIG_TCC_BCHECK
-    /* The main reason to removing this code:
-	#include <stdio.h>
-	int main ()
-	{
-	    int v[10];
-	    int i = 10;
-	    int j = 9;
-	    fprintf(stderr, "v+i-j  = %p\n", v+i-j);
-	    fprintf(stderr, "v+(i-j)  = %p\n", v+(i-j));
-	}
-    When this code is on. then the output looks like 
-	v+i-j = 0xfffffffe
-	v+(i-j) = 0xbff84000
-    This should now work in updated bcheck.c version.
-    */
-            /* if evaluating constant expression, no code should be
-               generated, so no bound check */
-            if (tcc_state->do_bounds_check && tcc_state->do_bounds_check_address
-                && !const_wanted && !nocode_wanted) {
+            if (tcc_state->do_bounds_check && !const_wanted) {
                 /* if bounded pointers, we generate a special code to
                    test bounds */
                 if (op == '-') {
@@ -2525,6 +2525,7 @@ redo:
                     vswap();
                     gen_op('-');
                 }
+                vtop[-1].r &= ~VT_MUSTBOUND;
                 gen_bounded_ptr_add();
             } else
 #endif
diff --git a/tccpe.c b/tccpe.c
index 6c15e048..01318598 100644
--- a/tccpe.c
+++ b/tccpe.c
@@ -1836,6 +1836,12 @@ ST_FUNC void pe_add_unwind_data(unsigned start, unsigned end, unsigned stack)
 #define PE_STDSYM(n,s) "_" n s
 #endif
 
+static void tcc_add_support(TCCState *s1, const char *filename)
+{
+    if (tcc_add_dll(s1, filename, 0) < 0)
+        tcc_error_noabort("%s not found", filename);
+}
+
 static void pe_add_runtime(TCCState *s1, struct pe_info *pe)
 {
     const char *start_symbol;
@@ -1884,35 +1890,21 @@ static void pe_add_runtime(TCCState *s1, struct pe_info *pe)
 
     if (0 == s1->nostdlib) {
         static const char *libs[] = {
-            TCC_LIBTCC1,
-#ifdef CONFIG_TCC_BCHECK
-            TCC_LIBTCCB1,
-#endif
             "msvcrt", "kernel32", "", "user32", "gdi32", NULL
         };
         const char **pp, *p;
+#ifdef TCC_IS_NATIVE
+        if (s1->do_bounds_check)
+            tcc_add_support(s1, "bcheck.o");
+#endif
+        tcc_add_support(s1, TCC_LIBTCC1);
         for (pp = libs; 0 != (p = *pp); ++pp) {
-#ifdef CONFIG_TCC_BCHECK
-            if (pp == libs + 1 &&
-                (s1->do_bounds_check == 0 || s1->output_type == TCC_OUTPUT_DLL)) {
-                continue;
-            }
-#endif
-            if (0 == *p) {
-                if (PE_DLL != pe_type && PE_GUI != pe_type)
-                    break;
-            } else if (pp == libs && tcc_add_dll(s1, p, 0) >= 0) {
-                continue;
-#ifdef CONFIG_TCC_BCHECK
-            } else if (pp == libs + 1 && tcc_add_dll(s1, p, 0) >= 0) {
-                continue;
-#endif
-            } else {
+            if (*p)
                 tcc_add_library_err(s1, p);
-            }
+            else if (PE_DLL != pe_type && PE_GUI != pe_type)
+                break;
         }
     }
-
     if (TCC_OUTPUT_MEMORY == s1->output_type)
         pe_type = PE_RUN;
     pe->type = pe_type;
diff --git a/tccpp.c b/tccpp.c
index 8245a8fd..5f66cce1 100644
--- a/tccpp.c
+++ b/tccpp.c
@@ -392,16 +392,15 @@ ST_FUNC void cstr_reset(CString *cstr)
 ST_FUNC int cstr_printf(CString *cstr, const char *fmt, ...)
 {
     va_list v;
-    va_list vc;
     int len, size;
 
     va_start(v, fmt);
-    va_copy (vc, v);
-    len = vsnprintf(NULL, 0, fmt, vc);
-    va_end(vc);
+    len = vsnprintf(NULL, 0, fmt, v);
+    va_end(v);
     size = cstr->size + len + 1;
     if (size > cstr->size_allocated)
         cstr_realloc(cstr, size);
+    va_start(v, fmt);
     vsnprintf((char*)cstr->data + cstr->size, size, fmt, v);
     va_end(v);
     cstr->size += len;
diff --git a/tcctok.h b/tcctok.h
index 6de1c6e9..5adafb9c 100644
--- a/tcctok.h
+++ b/tcctok.h
@@ -306,6 +306,9 @@
      DEF(TOK___bound_local_new, "__bound_local_new")
      DEF(TOK___bound_local_delete, "__bound_local_delete")
 # ifdef TCC_TARGET_PE
+#  ifdef TCC_TARGET_X86_64
+     DEF(TOK___bound_alloca_nr, "__bound_alloca_nr")
+#  endif
      DEF(TOK_malloc, "malloc")
      DEF(TOK_free, "free")
      DEF(TOK_realloc, "realloc")
diff --git a/tests/Makefile b/tests/Makefile
index e8507f58..ce434b10 100644
--- a/tests/Makefile
+++ b/tests/Makefile
@@ -23,7 +23,7 @@ TESTS = \
  tests2-dir \
  pp-dir
 
-BTESTS = test1b test3b btest test4
+BTESTS = btest test2b
 
 # test4_static -- Not all relocation types are implemented yet.
 # asmtest / asmtest2 -- minor differences with gcc
@@ -115,7 +115,7 @@ test3 test3b: tcctest.c test.ref
 	$(TCC) $(RUN_TCC) $(RUN_TCC) $(RUN_TCC) -run $< > test.out3
 	@diff -u test.ref test.out3 && echo "Auto Test3 OK"
 
-test%b : TCCFLAGS += -b -ba
+test%b : TCCFLAGS += -b
 
 # binary output test
 test4: tcctest.c test.ref
@@ -130,7 +130,7 @@ test4: tcctest.c test.ref
 	./tcctest1 > test1.out
 	@if diff -u test.ref test1.out ; then echo "Dynamic Auto Test OK"; fi
 # dynamic output + bound check
-	$(TCC) -b -ba -o tcctest4 $<
+	$(TCC) -b -o tcctest4 $<
 	./tcctest4 > test4.out
 	@if diff -u test.ref test4.out ; then echo "BCheck Auto Test OK"; fi
 
@@ -163,15 +163,14 @@ memtest:
 
 
 # memory and bound check auto test
-BOUNDS_OK  = 1 3 4 8 10 14 16
+BOUNDS_OK  = 1 4 8 10 14 16
 BOUNDS_FAIL= 2 5 6 7 9 11 12 13 15 17
 
 btest: boundtest.c
 	@echo ------------ $@ ------------
-	@ulimit -c 0; \
-	for i in $(BOUNDS_OK); do \
+	@for i in $(BOUNDS_OK); do \
 	   echo ; echo --- boundtest $$i ---; \
-	   if $(TCC) -b -ba -run $< $$i ; then \
+	   if $(TCC) -b -run $< $$i ; then \
 	       echo succeeded as expected; \
 	   else\
 	       echo Failed positive test $$i ; exit 1 ; \
@@ -179,7 +178,7 @@ btest: boundtest.c
 	done ;\
 	for i in $(BOUNDS_FAIL); do \
 	   echo ; echo --- boundtest $$i ---; \
-	   if $(TCC) -b -ba -run $< $$i ; then \
+	   if $(TCC) -b -run $< $$i ; then \
 	       echo Failed negative test $$i ; exit 1 ;\
 	   else\
 	       echo failed as expected; \
diff --git a/x86_64-gen.c b/x86_64-gen.c
index c1cce786..bdbd137d 100644
--- a/x86_64-gen.c
+++ b/x86_64-gen.c
@@ -142,9 +142,6 @@ ST_DATA const int reg_classes[NB_REGS] = {
 
 static unsigned long func_sub_sp_offset;
 static int func_ret_sub;
-static int nested_call;
-static int call_used_nr_reg;
-static int call_used_regs[20];
 
 /* XXX: make it faster ? */
 ST_FUNC void g(int c)
@@ -654,218 +651,19 @@ static void gen_bounds_call(int v)
 #endif
 }
 
-ST_FUNC void tcc_add_bcheck(TCCState *s1, Section *bound_sec, Section *sym_sec)
-{
-    addr_t *ptr;
-    int loc_glob;
-    int sym_index;
-    int bsym_index;
-
-    if (0 == s1->do_bounds_check)
-        return;
-    /* XXX: add an object file to do that */
-    ptr = section_ptr_add(bound_sec, sizeof(*ptr));
-    *ptr = 0;
-    loc_glob = s1->output_type != TCC_OUTPUT_MEMORY ? STB_LOCAL : STB_GLOBAL;
-    bsym_index = set_elf_sym(sym_sec, 0, 0,
-                ELFW(ST_INFO)(loc_glob, STT_NOTYPE), 0,
-                bound_sec->sh_num, "__bounds_start");
-    /* pull bcheck.o from libtcc1.a */
-    sym_index = set_elf_sym(sym_sec, 0, 0,
-                ELFW(ST_INFO)(STB_GLOBAL, STT_NOTYPE), 0,
-                SHN_UNDEF, "__bound_init");
-    if (s1->output_type != TCC_OUTPUT_MEMORY) {
-        /* add 'call __bound_init()' in .init section */
-        Section *init_section = find_section(s1, ".init");
-        unsigned char *pinit;
-#ifdef TCC_TARGET_PE
-        pinit = section_ptr_add(init_section, 8);
-        pinit[0] = 0x55;        /* push %rbp */
-        pinit[1] = 0x48;        /* mov %rsp,%rpb */
-        pinit[2] = 0x89;
-        pinit[3] = 0xe5;
-        pinit[4] = 0x48;        /* sub $0x10,%rsp */
-        pinit[5] = 0x83;
-        pinit[6] = 0xec;
-        pinit[7] = 0x10;
-#endif
-        pinit = section_ptr_add(init_section, 5);
-        pinit[0] = 0xe8;
-        write32le(pinit + 1, -4);
-        put_elf_reloc(sym_sec, init_section,
-            init_section->data_offset - 4, R_386_PC32, sym_index);
-            /* R_386_PC32 = R_X86_64_PC32 = 2 */
-        pinit = section_ptr_add(init_section, 13);
-        pinit[0] = 0x48;        /* mov xx,%rax */
-        pinit[1] = 0xb8;
-        write64le(pinit + 2, 0);
-#ifdef TCC_TARGET_PE
-        pinit[10] = 0x48;        /* mov %rax,%rcx */
-        pinit[11] = 0x89;
-        pinit[12] = 0xc1;
-#else
-        pinit[10] = 0x48;        /* mov %rax,%rdi */
-        pinit[11] = 0x89;
-        pinit[12] = 0xc7;
-#endif
-        put_elf_reloc(sym_sec, init_section,
-                init_section->data_offset - 11, R_X86_64_64, bsym_index);
-        sym_index = set_elf_sym(sym_sec, 0, 0,
-                        ELFW(ST_INFO)(STB_GLOBAL, STT_NOTYPE), 0,
-                        SHN_UNDEF, "__bounds_add_static_var");
-        pinit = section_ptr_add(init_section, 5);
-        pinit[0] = 0xe8;
-        write32le(pinit + 1, -4);
-        put_elf_reloc(sym_sec, init_section,
-            init_section->data_offset - 4, R_386_PC32, sym_index);
-                /* R_386_PC32 = R_X86_64_PC32 = 2 */
-#ifdef TCC_TARGET_PE
-        {
-            int init_index = set_elf_sym(sym_sec,
-                                         0, 0,
-                                         ELFW(ST_INFO)(STB_GLOBAL, STT_NOTYPE), 0,
-                                         init_section->sh_num, "__init_start");
-            Sym sym;
-            init_section->sh_flags |= SHF_EXECINSTR;
-            pinit = section_ptr_add(init_section, 2);
-            pinit[0] = 0xc9;        /* leave */
-            pinit[1] = 0xc3;        /* ret */
-            sym.c = init_index;
-            add_init_array (s1, &sym);
-        }
-#endif
-    }
-}
-
 /* generate a bounded pointer addition */
 ST_FUNC void gen_bounded_ptr_add(void)
 {
-    int i;
-
-    nested_call++;
-
-    /* save all temporary registers */
-    save_regs(0);
-
-    for (i = 0; i < call_used_nr_reg; i++) {
-        switch (call_used_regs[i]) {
-        case TREG_RAX: case TREG_RCX: case TREG_RDX: case TREG_RSP:
-        case TREG_RSI: case TREG_RDI:
-            o(0x50 + call_used_regs[i]);    /* push reg */
-            break;
-        case TREG_R8: case TREG_R9: case TREG_R10: case TREG_R11:
-            o(0x5041 + (call_used_regs[i] - TREG_R8) * 0x100);  /* push reg */
-            break;
-        case TREG_XMM0: case TREG_XMM1: case TREG_XMM2: case TREG_XMM3:
-        case TREG_XMM4: case TREG_XMM5: case TREG_XMM6: case TREG_XMM7:
-            o(0x10ec8348);             /* sub $10,%rsp */
-                                       /* vmovdqu %xmmx,(%rsp) */
-            o(0x047ffac5 + (call_used_regs[i] - TREG_XMM0) * 0x8000000); o(0x24);
-            break;
-        }
-    }
-
-    if (nested_call > 1) {
-#ifdef TCC_TARGET_PE
-        o(0x5152);       /* push %rdx/%rcx */
-        o(0x51415041);   /* push %r8/%r9 */
-#else
-        o(0x51525657);   /* push %rdi/%rsi/%rdx/%rcx */
-#endif
-    }
-    /* prepare fast x86_64 function call */
-    gv(RC_RAX);
-#ifdef TCC_TARGET_PE
-    o(0xc28948); // mov  %rax,%rdx ## second arg in %rdx, this must be size
-#else
-    o(0xc68948); // mov  %rax,%rsi ## second arg in %rsi, this must be size
-#endif
-    vtop--;
-
-    gv(RC_RAX);
-#ifdef TCC_TARGET_PE
-    o(0xc18948); // mov  %rax,%rcx ## first arg in %rcx, this must be ptr
-#else
-    o(0xc78948); // mov  %rax,%rdi ## first arg in %rdi, this must be ptr
-#endif
-    vtop--;
-
-    /* add line, filename */
-#ifdef TCC_TARGET_PE
-    o(0xc0c749);   /* mov $xx,%r8 */
-#else
-    o(0xba);       /* mov $xx,%edx */
-#endif
-    gen_le32 (file->line_num);
-    {
-        static addr_t offset;
-        static char last_filename[1024];
-        Sym *sym_data;
-
-        if (strcmp (last_filename, file->filename) != 0) {
-            void *ptr;
-            int len = strlen (file->filename) + 1;
-
-            offset = data_section->data_offset;
-            ptr = section_ptr_add(data_section, len);
-            memcpy (ptr, file->filename, len);
-            memcpy (last_filename, file->filename, len);
-        }
-#ifdef TCC_TARGET_PE
-        o(0xb949);   /* mov $xx,%r9 */
-#else
-        o(0xb948);   /* mov $xx,%rcx */
-#endif
-        gen_le64 (0);
-        sym_data = get_sym_ref(&char_pointer_type, data_section,
-                               offset, data_section->data_offset);
-        greloca(cur_text_section, sym_data, ind - 8, R_X86_64_64, 0);
-    }
-
-#ifdef TCC_TARGET_PE
-    o(0x20ec8348); /* sub $20, %rsp */
-#endif
-
-    /* do a fast function call */
-    gen_bounds_call(TOK___bound_ptr_add);
-
-#ifdef TCC_TARGET_PE
-    o(0x20c48348); /* add $20, %rsp */
-#endif
-
-    if (nested_call > 1) {
-#ifdef TCC_TARGET_PE
-        o(0x58415941);   /* pop %r9/%r8 */
-        o(0x5a59);       /* pop %rcx/%rdx */
-#else
-        o(0x5f5e5a59);   /* pop $rcx/%rdx/%rsi/%rdi */
-#endif
-    }
+    vpush_global_sym(&func_old_type, TOK___bound_ptr_add);
+    vrott(3);
+    gfunc_call(2);
+    vpushi(0);
     /* returned pointer is in rax */
-    vtop++;
     vtop->r = TREG_RAX | VT_BOUNDED;
-
-    for (i = call_used_nr_reg - 1; i >= 0; i--) {
-        switch (call_used_regs[i]) {
-        case TREG_RAX: case TREG_RCX: case TREG_RDX: case TREG_RSP:
-        case TREG_RSI: case TREG_RDI:
-            o(0x58 + call_used_regs[i]);    /* pop reg */
-            break;
-        case TREG_R8: case TREG_R9: case TREG_R10: case TREG_R11:
-            o(0x5841 + (call_used_regs[i] - TREG_R8) * 0x100);  /* pop reg */
-            break;
-        case TREG_XMM0: case TREG_XMM1: case TREG_XMM2: case TREG_XMM3:
-        case TREG_XMM4: case TREG_XMM5: case TREG_XMM6: case TREG_XMM7:
-                                       /* vmovdqu (%rsp),%xmmx */
-            o(0x046ffac5 + (call_used_regs[i] - TREG_XMM0) * 0x8000000); o(0x24);
-            o(0x10c48348);             /* add $10,%rsp */
-            break;
-        }
-    }
-
+    if (nocode_wanted)
+        return;
     /* relocation offset of the bounding function call point */
     vtop->c.i = (cur_text_section->reloc->data_offset - sizeof(ElfW(Rela)));
-    nested_call--;
 }
 
 /* patch pointer addition in vtop so that pointer dereferencing is
@@ -877,6 +675,9 @@ ST_FUNC void gen_bounded_ptr_deref(void)
     ElfW(Rela) *rel;
     Sym *sym;
 
+    if (nocode_wanted)
+        return;
+
     size = 0;
     /* XXX: put that code in generic part of tcc */
     if (!is_float(vtop->type.t)) {
@@ -886,7 +687,7 @@ ST_FUNC void gen_bounded_ptr_deref(void)
             size = 2;
     }
     if (!size)
-    size = type_size(&vtop->type, &align);
+        size = type_size(&vtop->type, &align);
     switch(size) {
     case  1: func = TOK___bound_ptr_indir1; break;
     case  2: func = TOK___bound_ptr_indir2; break;
@@ -895,18 +696,17 @@ ST_FUNC void gen_bounded_ptr_deref(void)
     case 12: func = TOK___bound_ptr_indir12; break;
     case 16: func = TOK___bound_ptr_indir16; break;
     default:
-        tcc_error("unhandled size when dereferencing bounded pointer");
-        func = 0;
-        break;
+        /* may happen with struct member access */
+        return;
+        //tcc_error("unhandled size when dereferencing bounded pointer");
+        //func = 0;
+        //break;
     }
-
     sym = external_global_sym(func, &func_old_type);
     if (!sym->c)
         put_extern_sym(sym, NULL, 0, 0);
-
     /* patch relocation */
     /* XXX: find a better solution ? */
-
     rel = (ElfW(Rela) *)(cur_text_section->reloc->data + vtop->c.i);
     rel->r_info = ELF64_R_INFO(sym->c, ELF64_R_TYPE(rel->r_info));
 }
@@ -992,6 +792,11 @@ void gfunc_call(int nb_args)
     int size, r, args_size, i, d, bt, struct_size;
     int arg;
 
+#ifdef CONFIG_TCC_BCHECK
+    if (tcc_state->do_bounds_check)
+        gbound_args(nb_args);
+#endif
+
     args_size = (nb_args < REGN ? REGN : nb_args) * PTR_SIZE;
     arg = nb_args;
 
@@ -1051,7 +856,6 @@ void gfunc_call(int nb_args)
             } else {
                 d = arg_prepare_reg(arg);
                 gen_offs_sp(0x8d, d, struct_size);
-                call_used_regs[call_used_nr_reg++] = d;
             }
             struct_size += size;
         } else {
@@ -1070,7 +874,6 @@ void gfunc_call(int nb_args)
                     o(0x66);
                     orex(1,d,0, 0x7e0f);
                     o(0xc0 + arg*8 + REG_VALUE(d));
-                    call_used_regs[call_used_nr_reg++] = d;
                 }
             } else {
                 if (bt == VT_STRUCT) {
@@ -1086,7 +889,6 @@ void gfunc_call(int nb_args)
                     d = arg_prepare_reg(arg);
                     orex(1,d,r,0x89); /* mov */
                     o(0xc0 + REG_VALUE(r) * 8 + REG_VALUE(d));
-                    call_used_regs[call_used_nr_reg++] = d;
                 }
             }
         }
@@ -1105,7 +907,12 @@ void gfunc_call(int nb_args)
 
     if ((vtop->r & VT_SYM) && vtop->sym->v == TOK_alloca) {
         /* need to add the "func_scratch" area after alloca */
-        o(0x48); func_alloca = oad(0x2d, func_alloca); /* sub $NN, %rax */
+        o(0x48); func_alloca = oad(0x05, func_alloca); /* add $NN, %rax */
+#ifdef CONFIG_TCC_BCHECK
+        if (tcc_state->do_bounds_check)
+            gen_bounds_call(TOK___bound_alloca_nr); /* new region */
+#endif
+
     }
 
     /* other compilers don't clear the upper bits when returning char/short */
@@ -1125,7 +932,6 @@ void gfunc_call(int nb_args)
         o(0xc089); /* mov %eax,%eax */
 #endif
     vtop--;
-    call_used_nr_reg = 0;
 }
 
 
@@ -1141,7 +947,7 @@ void gfunc_prolog(Sym *func_sym)
     int n_arg = 0;
 
     func_ret_sub = 0;
-    func_scratch = 0;
+    func_scratch = 32;
     func_alloca = 0;
     loc = 0;
 
@@ -1228,6 +1034,10 @@ void gfunc_epilog(void)
 {
     int v, saved_ind;
 
+    /* align local size to word & save local variables */
+    func_scratch = (func_scratch + 15) & -16;
+    loc = (loc & -16) - func_scratch;
+
 #ifdef CONFIG_TCC_BCHECK
     if (tcc_state->do_bounds_check
         && func_bound_offset != lbounds_section->data_offset)
@@ -1273,9 +1083,7 @@ void gfunc_epilog(void)
 
     saved_ind = ind;
     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
-    /* align local size to word & save local variables */
-    func_scratch = (func_scratch + 15) & -16;
-    v = (func_scratch + -loc + 15) & -16;
+    v = -loc;
 
     if (v >= 4096) {
         Sym *sym = external_global_sym(TOK___chkstk, &func_old_type);
@@ -1479,6 +1287,11 @@ void gfunc_call(int nb_args)
     int sse_reg, gen_reg;
     char _onstack[nb_args ? nb_args : 1], *onstack = _onstack;
 
+#ifdef CONFIG_TCC_BCHECK
+    if (tcc_state->do_bounds_check)
+        gbound_args(nb_args);
+#endif
+
     /* calculate the number of integer/float register arguments, remember
        arguments to be passed via stack (in onstack[]), and also remember
        if we have to align the stack pointer to 16 (onstack[i] == 2).  Needs
@@ -1611,14 +1424,11 @@ void gfunc_call(int nb_args)
                     o(0x280f);
                     o(0xc0 + (sse_reg << 3));
                 }
-                call_used_regs[call_used_nr_reg++] = sse_reg + TREG_XMM0;
-                call_used_regs[call_used_nr_reg++] = sse_reg + 1 + TREG_XMM0;
             } else {
                 assert(reg_count == 1);
                 --sse_reg;
                 /* Load directly to register */
                 gv(RC_XMM0 << sse_reg);
-                call_used_regs[call_used_nr_reg++] = sse_reg + TREG_XMM0;
             }
         } else if (mode == x86_64_mode_integer) {
             /* simple type */
@@ -1629,12 +1439,10 @@ void gfunc_call(int nb_args)
             d = arg_prepare_reg(gen_reg);
             orex(1,d,r,0x89); /* mov */
             o(0xc0 + REG_VALUE(r) * 8 + REG_VALUE(d));
-            call_used_regs[call_used_nr_reg++] = d;
             if (reg_count == 2) {
                 d = arg_prepare_reg(gen_reg+1);
                 orex(1,d,vtop->r2,0x89); /* mov */
                 o(0xc0 + REG_VALUE(vtop->r2) * 8 + REG_VALUE(d));
-                call_used_regs[call_used_nr_reg++] = d;
             }
         }
         vtop--;
@@ -1677,7 +1485,6 @@ void gfunc_call(int nb_args)
     else if (bt == (VT_SHORT | VT_UNSIGNED))
         o(0xc0b70f);  /* movzwl %al, %eax */
     vtop--;
-    call_used_nr_reg = 0;
 }
 
 
diff --git a/x86_64-link.c b/x86_64-link.c
index dcb2b90c..0730c2db 100644
--- a/x86_64-link.c
+++ b/x86_64-link.c
@@ -287,4 +287,89 @@ void relocate(TCCState *s1, ElfW_Rel *rel, int type, unsigned char *ptr, addr_t
     }
 }
 
+#ifdef CONFIG_TCC_BCHECK
+ST_FUNC void tcc_add_bcheck(TCCState *s1)
+{
+    addr_t *ptr;
+    int loc_glob;
+    int sym_index;
+    int bsym_index;
+
+    if (0 == s1->do_bounds_check)
+        return;
+    /* XXX: add an object file to do that */
+    ptr = section_ptr_add(bounds_section, sizeof(*ptr));
+    *ptr = 0;
+    loc_glob = s1->output_type != TCC_OUTPUT_MEMORY ? STB_LOCAL : STB_GLOBAL;
+    bsym_index = set_elf_sym(symtab_section, 0, 0,
+                ELFW(ST_INFO)(loc_glob, STT_NOTYPE), 0,
+                bounds_section->sh_num, "__bounds_start");
+    /* pull bcheck.o from libtcc1.a */
+    sym_index = set_elf_sym(symtab_section, 0, 0,
+                ELFW(ST_INFO)(STB_GLOBAL, STT_NOTYPE), 0,
+                SHN_UNDEF, "__bound_init");
+    if (s1->output_type != TCC_OUTPUT_MEMORY) {
+        /* add 'call __bound_init()' in .init section */
+        Section *init_section = find_section(s1, ".init");
+        unsigned char *pinit;
+#ifdef TCC_TARGET_PE
+        pinit = section_ptr_add(init_section, 8);
+        pinit[0] = 0x55;        /* push %rbp */
+        pinit[1] = 0x48;        /* mov %rsp,%rpb */
+        pinit[2] = 0x89;
+        pinit[3] = 0xe5;
+        pinit[4] = 0x48;        /* sub $0x10,%rsp */
+        pinit[5] = 0x83;
+        pinit[6] = 0xec;
+        pinit[7] = 0x10;
+#endif
+        pinit = section_ptr_add(init_section, 5);
+        pinit[0] = 0xe8;
+        write32le(pinit + 1, -4);
+        put_elf_reloc(symtab_section, init_section,
+            init_section->data_offset - 4, R_386_PC32, sym_index);
+            /* R_386_PC32 = R_X86_64_PC32 = 2 */
+        pinit = section_ptr_add(init_section, 13);
+        pinit[0] = 0x48;        /* mov xx,%rax */
+        pinit[1] = 0xb8;
+        write64le(pinit + 2, 0);
+#ifdef TCC_TARGET_PE
+        pinit[10] = 0x48;        /* mov %rax,%rcx */
+        pinit[11] = 0x89;
+        pinit[12] = 0xc1;
+#else
+        pinit[10] = 0x48;        /* mov %rax,%rdi */
+        pinit[11] = 0x89;
+        pinit[12] = 0xc7;
+#endif
+        put_elf_reloc(symtab_section, init_section,
+                init_section->data_offset - 11, R_X86_64_64, bsym_index);
+        sym_index = set_elf_sym(symtab_section, 0, 0,
+                        ELFW(ST_INFO)(STB_GLOBAL, STT_NOTYPE), 0,
+                        SHN_UNDEF, "__bounds_add_static_var");
+        pinit = section_ptr_add(init_section, 5);
+        pinit[0] = 0xe8;
+        write32le(pinit + 1, -4);
+        put_elf_reloc(symtab_section, init_section,
+            init_section->data_offset - 4, R_386_PC32, sym_index);
+                /* R_386_PC32 = R_X86_64_PC32 = 2 */
+#ifdef TCC_TARGET_PE
+        {
+            int init_index = set_elf_sym(symtab_section,
+                                         0, 0,
+                                         ELFW(ST_INFO)(STB_GLOBAL, STT_NOTYPE), 0,
+                                         init_section->sh_num, "__init_start");
+            Sym sym;
+            init_section->sh_flags |= SHF_EXECINSTR;
+            pinit = section_ptr_add(init_section, 2);
+            pinit[0] = 0xc9;        /* leave */
+            pinit[1] = 0xc3;        /* ret */
+            sym.c = init_index;
+            add_init_array (s1, &sym);
+        }
+#endif
+    }
+}
+#endif
+
 #endif /* !TARGET_DEFS_ONLY */