win32: enable bounds checker & exception handler

exception handler borrowed from k1w1. Thanks.

11 files changed, 421 insertions, 331 deletions

diff --git a/Makefile b/Makefile
index e6dba91..08ef48a 100644
--- a/Makefile
+++ b/Makefile
@@ -78,7 +78,7 @@ ARM_CROSS = arm-tcc-fpa$(EXESUF) arm-tcc-fpa-ld$(EXESUF) \
     arm-tcc-vfp$(EXESUF) arm-tcc-vfp-eabi$(EXESUF)
 C67_CROSS = c67-tcc$(EXESUF)
 
-CORE_FILES = tcc.c libtcc.c tccpp.c tccgen.c tccelf.c tccasm.c \
+CORE_FILES = tcc.c libtcc.c tccpp.c tccgen.c tccelf.c tccasm.c tccrun.c \
     tcc.h config.h libtcc.h tcctok.h
 I386_FILES = $(CORE_FILES) i386-gen.c i386-asm.c i386-asm.h i386-tok.h
 WIN32_FILES = $(CORE_FILES) i386-gen.c i386-asm.c i386-asm.h i386-tok.h tccpe.c
@@ -180,7 +180,7 @@ VPATH+=lib
 
 ifdef CONFIG_WIN32
 # for windows, we must use TCC because we generate ELF objects
-LIBTCC1_OBJS+=crt1.o wincrt1.o dllcrt1.o dllmain.o chkstk.o
+LIBTCC1_OBJS+=crt1.o wincrt1.o dllcrt1.o dllmain.o chkstk.o bcheck.o
 LIBTCC1_CC=./tcc.exe -Bwin32 -Iinclude $(NATIVE_TARGET)
 VPATH+=win32/lib
 endif
@@ -193,9 +193,6 @@ endif
 libtcc1.a: $(LIBTCC1_OBJS)
 	$(AR) rcs $@ $^
 
-bcheck.o: bcheck.c
-	$(CC) -o $@ -c $< -O2 -Wall
-
 # install
 TCC_INCLUDES = stdarg.h stddef.h stdbool.h float.h varargs.h tcclib.h
 INSTALL=install
diff --git a/i386-gen.c b/i386-gen.c
index bd53303..161be93 100644
--- a/i386-gen.c
+++ b/i386-gen.c
@@ -87,8 +87,10 @@ const int reg_classes[NB_REGS] = {
 /******************************************************/
 
 static unsigned long func_sub_sp_offset;
-static unsigned long func_bound_offset;
 static int func_ret_sub;
+#ifdef CONFIG_TCC_BCHECK
+static unsigned long func_bound_offset;
+#endif
 
 /* XXX: make it faster ? */
 void g(int c)
@@ -511,12 +513,14 @@ void gfunc_prolog(CType *func_type)
         func_ret_sub = 4;
 #endif
 
+#ifdef CONFIG_TCC_BCHECK
     /* leave some room for bound checking code */
     if (tcc_state->do_bounds_check) {
         oad(0xb8, 0); /* lbound section pointer */
         oad(0xb8, 0); /* call to function */
         func_bound_offset = lbounds_section->data_offset;
     }
+#endif
 }
 
 /* generate function epilog */
diff --git a/lib/bcheck.c b/lib/bcheck.c
index 90590f5..09f2a3a 100644
--- a/lib/bcheck.c
+++ b/lib/bcheck.c
@@ -33,12 +33,11 @@
 
 /* use malloc hooks. Currently the code cannot be reliable if no hooks */
 #define CONFIG_TCC_MALLOC_HOOKS
-
 #define HAVE_MEMALIGN
 
 #if defined(__FreeBSD__) || defined(__DragonFly__) || defined(__dietlibc__) \
-    || defined(__UCLIBC__) || defined(__OpenBSD__)
-#warning Bound checking not fully supported in this environment.
+    || defined(__UCLIBC__) || defined(__OpenBSD__) || defined(_WIN32)
+#warning Bound checking does not support malloc (etc.) in this environment.
 #undef CONFIG_TCC_MALLOC_HOOKS
 #undef HAVE_MEMALIGN
 #endif
@@ -152,9 +151,6 @@ static void bound_alloc_error(void)
     bound_error("not enough memory for bound checking code");
 }
 
-/* currently, tcc cannot compile that because we use GNUC extensions */
-#if !defined(__TINYC__)
-
 /* 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, int offset)
@@ -203,6 +199,13 @@ void * FASTCALL __bound_ptr_indir ## dsize (void *p, int offset)        \
     return p + offset;                                                  \
 }
 
+BOUND_PTR_INDIR(1)
+BOUND_PTR_INDIR(2)
+BOUND_PTR_INDIR(4)
+BOUND_PTR_INDIR(8)
+BOUND_PTR_INDIR(12)
+BOUND_PTR_INDIR(16)
+
 #ifdef __i386__
 /* return the frame pointer of the caller */
 #define GET_CALLER_FP(fp)\
@@ -246,34 +249,6 @@ void FASTCALL __bound_local_delete(void *p1)
     }
 }
 
-#else
-
-void __bound_local_new(void *p) 
-{
-}
-void __bound_local_delete(void *p) 
-{
-}
-
-void *__bound_ptr_add(void *p, int offset)
-{
-    return p + offset;
-}
-
-#define BOUND_PTR_INDIR(dsize)                               \
-void *__bound_ptr_indir ## dsize (void *p, int offset)       \
-{                                                            \
-    return p + offset;                                       \
-}
-#endif
-
-BOUND_PTR_INDIR(1)
-BOUND_PTR_INDIR(2)
-BOUND_PTR_INDIR(4)
-BOUND_PTR_INDIR(8)
-BOUND_PTR_INDIR(12)
-BOUND_PTR_INDIR(16)
-
 static BoundEntry *__bound_new_page(void)
 {
     BoundEntry *page;
diff --git a/libtcc.c b/libtcc.c
index 9309b77..7ef1f0b 100644
--- a/libtcc.c
+++ b/libtcc.c
@@ -58,9 +58,13 @@ static Section *cur_text_section; /* current section where function code is
 #ifdef CONFIG_TCC_ASM
 static Section *last_text_section; /* to handle .previous asm directive */
 #endif
+
+#ifdef CONFIG_TCC_BCHECK
 /* bound check related sections */
 static Section *bounds_section; /* contains global data bound description */
 static Section *lbounds_section; /* contains local data bound description */
+#endif
+
 /* symbol sections */
 static Section *symtab_section, *strtab_section;
 
@@ -109,8 +113,8 @@ static int tcc_ext = 1;
 /* max number of callers shown if error */
 #ifdef CONFIG_TCC_BACKTRACE
 int num_callers = 6;
+void *rt_prog_main;
 const char **rt_bound_error_msg;
-unsigned long rt_prog_main;
 #endif
 
 /* XXX: get rid of this ASAP */
@@ -731,7 +735,7 @@ static void put_extern_sym2(Sym *sym, Section *section,
             /* if bound checking is activated, we change some function
                names by adding the "__bound" prefix */
             switch(sym->v) {
-#if 0
+#ifdef TCC_TARGET_PE
             /* XXX: we rely only on malloc hooks */
             case TOK_malloc: 
             case TOK_free: 
diff --git a/tcc.h b/tcc.h
index cc63e92..16aba53 100644
--- a/tcc.h
+++ b/tcc.h
@@ -101,7 +101,7 @@
 #define TCC_TARGET_I386
 #endif
 
-#if !defined(_WIN32) && !defined(TCC_UCLIBC) && !defined(TCC_TARGET_ARM) && \
+#if !defined(TCC_UCLIBC) && !defined(TCC_TARGET_ARM) && \
     !defined(TCC_TARGET_C67) && !defined(TCC_TARGET_X86_64)
 #define CONFIG_TCC_BCHECK /* enable bound checking code */
 #endif
@@ -120,7 +120,7 @@
 #define TCC_TARGET_COFF
 #endif
 
-#if !defined(_WIN32) && !defined(CONFIG_TCCBOOT)
+#if !defined(CONFIG_TCCBOOT)
 #define CONFIG_TCC_BACKTRACE
 #endif
 
@@ -465,8 +465,10 @@ struct TCCState {
     int verbose;
     /* compile with debug symbol (and use them if error during execution) */
     int do_debug;
+#ifdef CONFIG_TCC_BCHECK
     /* compile with built-in memory and bounds checker */
     int do_bounds_check;
+#endif
     /* give the path of the tcc libraries */
     char *tcc_lib_path;
 
@@ -809,11 +811,6 @@ char *pstrcat(char *buf, int buf_size, const char *s);
 void dynarray_add(void ***ptab, int *nb_ptr, void *data);
 void dynarray_reset(void *pp, int *n);
 
-#ifdef CONFIG_TCC_BACKTRACE
-extern int num_callers;
-extern const char **rt_bound_error_msg;
-#endif
-
 /* true if float/double/long double type */
 static inline int is_float(int t)
 {
diff --git a/tccelf.c b/tccelf.c
index 7b891cf..1b4a804 100644
--- a/tccelf.c
+++ b/tccelf.c
@@ -1186,50 +1186,59 @@ static void add_init_array_defines(TCCState *s1, const char *section_name)
                 s->sh_num, sym_end);
 }
 
-/* add tcc runtime libraries */
-static void tcc_add_runtime(TCCState *s1)
+static void tcc_add_bcheck(TCCState *s1)
 {
-#if defined(CONFIG_TCC_BCHECK) || !defined(CONFIG_USE_LIBGCC)
-    char buf[1024];
-#endif
-
 #ifdef CONFIG_TCC_BCHECK
-    if (s1->do_bounds_check) {
-        unsigned long *ptr;
-        Section *init_section;
-        unsigned char *pinit;
-        int sym_index;
+    unsigned long *ptr;
+    Section *init_section;
+    unsigned char *pinit;
+    int sym_index;
 
-        /* XXX: add an object file to do that */
-        ptr = section_ptr_add(bounds_section, sizeof(unsigned long));
-        *ptr = 0;
-        add_elf_sym(symtab_section, 0, 0, 
-                    ELFW(ST_INFO)(STB_GLOBAL, STT_NOTYPE), 0,
-                    bounds_section->sh_num, "__bounds_start");
-        /* add bound check code */
-        snprintf(buf, sizeof(buf), "%s/%s", s1->tcc_lib_path, "bcheck.o");
-        tcc_add_file(s1, buf);
-#ifdef TCC_TARGET_I386
-        if (s1->output_type != TCC_OUTPUT_MEMORY) {
-            /* add 'call __bound_init()' in .init section */
-            init_section = find_section(s1, ".init");
-            pinit = section_ptr_add(init_section, 5);
-            pinit[0] = 0xe8;
-            put32(pinit + 1, -4);
-            sym_index = find_elf_sym(symtab_section, "__bound_init");
-            put_elf_reloc(symtab_section, init_section, 
-                          init_section->data_offset - 4, R_386_PC32, sym_index);
-        }
+    if (0 == s1->do_bounds_check)
+        return;
+
+    /* XXX: add an object file to do that */
+    ptr = section_ptr_add(bounds_section, sizeof(unsigned long));
+    *ptr = 0;
+    add_elf_sym(symtab_section, 0, 0,
+                ELFW(ST_INFO)(STB_GLOBAL, STT_NOTYPE), 0,
+                bounds_section->sh_num, "__bounds_start");
+    /* add bound check code */
+#ifndef TCC_TARGET_PE
+    {
+    char buf[1024];
+    snprintf(buf, sizeof(buf), "%s/%s", s1->tcc_lib_path, "bcheck.o");
+    tcc_add_file(s1, buf);
+    }
 #endif
+#ifdef TCC_TARGET_I386
+    if (s1->output_type != TCC_OUTPUT_MEMORY) {
+        /* add 'call __bound_init()' in .init section */
+        init_section = find_section(s1, ".init");
+        pinit = section_ptr_add(init_section, 5);
+        pinit[0] = 0xe8;
+        put32(pinit + 1, -4);
+        sym_index = find_elf_sym(symtab_section, "__bound_init");
+        put_elf_reloc(symtab_section, init_section,
+                      init_section->data_offset - 4, R_386_PC32, sym_index);
     }
 #endif
+#endif
+}
+
+/* add tcc runtime libraries */
+static void tcc_add_runtime(TCCState *s1)
+{
+    tcc_add_bcheck(s1);
+
     /* add libc */
     if (!s1->nostdlib) {
-        tcc_add_library(s1, "c");
-
 #ifdef CONFIG_USE_LIBGCC
+        tcc_add_library(s1, "c");
         tcc_add_file(s1, CONFIG_SYSROOT "/lib/libgcc_s.so.1");
 #else
+        char buf[1024];
+        tcc_add_library(s1, "c");
         snprintf(buf, sizeof(buf), "%s/%s", s1->tcc_lib_path, "libtcc1.a");
         tcc_add_file(s1, buf);
 #endif
diff --git a/tccgen.c b/tccgen.c
index 6bc8998..9866c55 100644
--- a/tccgen.c
+++ b/tccgen.c
@@ -2911,8 +2911,10 @@ static void indir(void)
         && (vtop->type.t & VT_BTYPE) != VT_FUNC) {
         vtop->r |= lvalue_type(vtop->type.t);
         /* if bound checking, the referenced pointer must be checked */
+#ifdef CONFIG_TCC_BCHECK
         if (tcc_state->do_bounds_check)
             vtop->r |= VT_MUSTBOUND;
+#endif
     }
 }
 
@@ -3328,9 +3330,11 @@ static void unary(void)
             /* an array is never an lvalue */
             if (!(vtop->type.t & VT_ARRAY)) {
                 vtop->r |= lvalue_type(vtop->type.t);
+#ifdef CONFIG_TCC_BCHECK
                 /* if bound checking, the referenced pointer must be checked */
                 if (tcc_state->do_bounds_check)
                     vtop->r |= VT_MUSTBOUND;
+#endif
             }
             next();
         } else if (tok == '[') {
@@ -4702,10 +4706,13 @@ static void decl_initializer_alloc(CType *type, AttributeDef *ad, int r,
     }
     if ((r & VT_VALMASK) == VT_LOCAL) {
         sec = NULL;
+#ifdef CONFIG_TCC_BCHECK
         if (tcc_state->do_bounds_check && (type->t & VT_ARRAY))
             loc--;
+#endif
         loc = (loc - size) & -align;
         addr = loc;
+#ifdef CONFIG_TCC_BCHECK
         /* handles bounds */
         /* XXX: currently, since we do only one pass, we cannot track
            '&' operators, so we add only arrays */
@@ -4718,6 +4725,7 @@ static void decl_initializer_alloc(CType *type, AttributeDef *ad, int r,
             bounds_ptr[0] = addr;
             bounds_ptr[1] = size;
         }
+#endif
         if (v) {
             /* local variable */
             sym_push(v, type, r, addr);
@@ -4774,9 +4782,11 @@ static void decl_initializer_alloc(CType *type, AttributeDef *ad, int r,
             /* very important to increment global pointer at this time
                because initializers themselves can create new initializers */
             data_offset += size;
+#ifdef CONFIG_TCC_BCHECK
             /* add padding if bound check */
             if (tcc_state->do_bounds_check)
                 data_offset++;
+#endif
             sec->data_offset = data_offset;
             /* allocate section space to put the data */
             if (sec->sh_type != SHT_NOBITS && 
@@ -4813,7 +4823,7 @@ static void decl_initializer_alloc(CType *type, AttributeDef *ad, int r,
             vsetc(type, VT_CONST | VT_SYM, &cval);
             vtop->sym = sym;
         }
-
+#ifdef CONFIG_TCC_BCHECK
         /* handles bounds now because the symbol must be defined
            before for the relocation */
         if (tcc_state->do_bounds_check) {
@@ -4825,6 +4835,7 @@ static void decl_initializer_alloc(CType *type, AttributeDef *ad, int r,
             bounds_ptr[0] = 0; /* relocated */
             bounds_ptr[1] = size;
         }
+#endif
     }
     if (has_init) {
         decl_initializer(type, sec, addr, 1, 0);
diff --git a/tccpe.c b/tccpe.c
index 69161f9..d603a5e 100644
--- a/tccpe.c
+++ b/tccpe.c
@@ -1700,15 +1700,15 @@ PUB_FN int pe_output_file(TCCState * s1, const char *filename)
     pe.filename = filename;
     pe.s1 = s1;
 
+    tcc_add_bcheck(s1);
     pe_add_runtime_ex(s1, &pe);
     relocate_common_syms(); /* assign bss adresses */
     tcc_add_linker_symbols(s1);
 
     ret = pe_check_symbols(&pe);
     if (ret)
-        return ret;
-
-    if (filename) {
+        ;
+    else if (filename) {
         if (PE_DLL == pe.type) {
             pe.reloc = new_section(pe.s1, ".reloc", SHT_PROGBITS, 0);
             /* XXX: check if is correct for arm-pe target */
diff --git a/tccrun.c b/tccrun.c
index d62e92a..e984c0f 100644
--- a/tccrun.c
+++ b/tccrun.c
@@ -20,75 +20,167 @@
 
 /* tccrun.c - support for tcc -run */
 
+#ifdef _WIN32
+#define ucontext_t CONTEXT
+#endif
 
-/********************************************************/
-#ifdef CONFIG_TCC_STATIC
+static void set_pages_executable(void *ptr, unsigned long length);
+static void set_exception_handler(void);
+static int rt_get_caller_pc(unsigned long *paddr, ucontext_t *uc, int level);
+static void rt_error(ucontext_t *uc, const char *fmt, ...);
+static int tcc_relocate_ex(TCCState *s1, void *ptr);
 
-#define RTLD_LAZY       0x001
-#define RTLD_NOW        0x002
-#define RTLD_GLOBAL     0x100
-#define RTLD_DEFAULT    NULL
+/* ------------------------------------------------------------- */
+/* Do all relocations (needed before using tcc_get_symbol())
+   Returns -1 on error. */
 
-/* dummy function for profiling */
-void *dlopen(const char *filename, int flag)
+int tcc_relocate(TCCState *s1)
 {
-    return NULL;
-}
+    int ret;
 
-void dlclose(void *p)
-{
+    ret = tcc_relocate_ex(s1, NULL);
+    if (-1 != ret) {
+        s1->runtime_mem = tcc_malloc(ret);
+        ret = tcc_relocate_ex(s1, s1->runtime_mem);
+    }
+    return ret;
 }
 
-const char *dlerror(void)
+/* launch the compiled program with the given arguments */
+int tcc_run(TCCState *s1, int argc, char **argv)
 {
-    return "error";
-}
+    int (*prog_main)(int, char **);
 
-typedef struct TCCSyms {
-    char *str;
-    void *ptr;
-} TCCSyms;
+    if (tcc_relocate(s1) < 0)
+        return -1;
 
-#define TCCSYM(a) { #a, &a, },
+    prog_main = tcc_get_symbol_err(s1, "main");
 
-/* add the symbol you want here if no dynamic linking is done */
-static TCCSyms tcc_syms[] = {
-#if !defined(CONFIG_TCCBOOT)
-    TCCSYM(printf)
-    TCCSYM(fprintf)
-    TCCSYM(fopen)
-    TCCSYM(fclose)
+#ifdef CONFIG_TCC_BACKTRACE
+    if (s1->do_debug)
+        set_exception_handler();
 #endif
-    { NULL, NULL },
-};
 
-void *resolve_sym(TCCState *s1, const char *symbol)
-{
-    TCCSyms *p;
-    p = tcc_syms;
-    while (p->str != NULL) {
-        if (!strcmp(p->str, symbol))
-            return p->ptr;
-        p++;
+#ifdef CONFIG_TCC_BCHECK
+    if (s1->do_bounds_check) {
+        void (*bound_init)(void);
+        void (*bound_exit)(void);
+        int ret;
+        /* set error function */
+        rt_bound_error_msg = tcc_get_symbol_err(s1, "__bound_error_msg");
+        rt_prog_main = prog_main;
+        /* XXX: use .init section so that it also work in binary ? */
+        bound_init = tcc_get_symbol_err(s1, "__bound_init");
+        bound_exit = tcc_get_symbol_err(s1, "__bound_exit");
+        bound_init();
+        ret = (*prog_main)(argc, argv);
+        bound_exit();
+        return ret;
     }
-    return NULL;
+#endif
+    return (*prog_main)(argc, argv);
 }
 
-#elif defined(_WIN32)
-#define dlclose FreeLibrary
 
+/* relocate code. Return -1 on error, required size if ptr is NULL,
+   otherwise copy code into buffer passed by the caller */
+static int tcc_relocate_ex(TCCState *s1, void *ptr)
+{
+    Section *s;
+    unsigned long offset, length;
+    uplong mem;
+    int i;
+
+    if (0 == s1->runtime_added) {
+        s1->runtime_added = 1;
+        s1->nb_errors = 0;
+#ifdef TCC_TARGET_PE
+        pe_output_file(s1, NULL);
 #else
-#include <dlfcn.h>
+        tcc_add_runtime(s1);
+        relocate_common_syms();
+        tcc_add_linker_symbols(s1);
+        build_got_entries(s1);
+#endif
+        if (s1->nb_errors)
+            return -1;
+    }
 
-void *resolve_sym(TCCState *s1, const char *sym)
-{
-    return dlsym(RTLD_DEFAULT, sym);
+    offset = 0, mem = (uplong)ptr;
+    for(i = 1; i < s1->nb_sections; i++) {
+        s = s1->sections[i];
+        if (0 == (s->sh_flags & SHF_ALLOC))
+            continue;
+        length = s->data_offset;
+        s->sh_addr = mem ? (mem + offset + 15) & ~15 : 0;
+        offset = (offset + length + 15) & ~15;
+    }
+    offset += 16;
+
+    /* relocate symbols */
+    relocate_syms(s1, 1);
+    if (s1->nb_errors)
+        return -1;
+
+#if defined TCC_TARGET_X86_64 && !defined TCC_TARGET_PE
+    s1->runtime_plt_and_got_offset = 0;
+    s1->runtime_plt_and_got = (char *)(mem + offset);
+    /* double the size of the buffer for got and plt entries
+       XXX: calculate exact size for them? */
+    offset *= 2;
+#endif
+
+    if (0 == mem)
+        return offset;
+
+    /* relocate each section */
+    for(i = 1; i < s1->nb_sections; i++) {
+        s = s1->sections[i];
+        if (s->reloc)
+            relocate_section(s1, s);
+    }
+
+    for(i = 1; i < s1->nb_sections; i++) {
+        s = s1->sections[i];
+        if (0 == (s->sh_flags & SHF_ALLOC))
+            continue;
+        length = s->data_offset;
+        // printf("%-12s %08x %04x\n", s->name, s->sh_addr, length);
+        ptr = (void*)(uplong)s->sh_addr;
+        if (NULL == s->data || s->sh_type == SHT_NOBITS)
+            memset(ptr, 0, length);
+        else
+            memcpy(ptr, s->data, length);
+        /* mark executable sections as executable in memory */
+        if (s->sh_flags & SHF_EXECINSTR)
+            set_pages_executable(ptr, length);
+    }
+
+#if defined TCC_TARGET_X86_64 && !defined TCC_TARGET_PE
+    set_pages_executable(s1->runtime_plt_and_got,
+                         s1->runtime_plt_and_got_offset);
+#endif
+    return 0;
 }
 
-#endif /* defined CONFIG_TCC_STATIC */
+/* ------------------------------------------------------------- */
+/* allow to run code in memory */
 
-/********************************************************/
+static void set_pages_executable(void *ptr, unsigned long length)
+{
+#ifdef _WIN32
+    unsigned long old_protect;
+    VirtualProtect(ptr, length, PAGE_EXECUTE_READWRITE, &old_protect);
+#else
+    unsigned long start, end;
+    start = (unsigned long)ptr & ~(PAGESIZE - 1);
+    end = (unsigned long)ptr + length;
+    end = (end + PAGESIZE - 1) & ~(PAGESIZE - 1);
+    mprotect((void *)start, end - start, PROT_READ | PROT_WRITE | PROT_EXEC);
+#endif
+}
 
+/* ------------------------------------------------------------- */
 #ifdef CONFIG_TCC_BACKTRACE
 
 /* print the position in the source file of PC value 'pc' by reading
@@ -215,7 +307,92 @@ static unsigned long rt_printline(unsigned long wanted_pc)
     return func_addr;
 }
 
+/* emit a run time error at position 'pc' */
+static void rt_error(ucontext_t *uc, const char *fmt, ...)
+{
+    va_list ap;
+    unsigned long pc;
+    int i;
+
+    va_start(ap, fmt);
+    fprintf(stderr, "Runtime error: ");
+    vfprintf(stderr, fmt, ap);
+    fprintf(stderr, "\n");
+
+    for(i=0;i<num_callers;i++) {
+        if (rt_get_caller_pc(&pc, uc, i) < 0)
+            break;
+        if (i == 0)
+            fprintf(stderr, "at ");
+        else
+            fprintf(stderr, "by ");
+        pc = rt_printline(pc);
+        if (pc == (unsigned long)rt_prog_main && pc)
+            break;
+    }
+    exit(255);
+    va_end(ap);
+}
+
+/* ------------------------------------------------------------- */
+#ifndef _WIN32
+
+/* signal handler for fatal errors */
+static void sig_error(int signum, siginfo_t *siginf, void *puc)
+{
+    ucontext_t *uc = puc;
+
+    switch(signum) {
+    case SIGFPE:
+        switch(siginf->si_code) {
+        case FPE_INTDIV:
+        case FPE_FLTDIV:
+            rt_error(uc, "division by zero");
+            break;
+        default:
+            rt_error(uc, "floating point exception");
+            break;
+        }
+        break;
+    case SIGBUS:
+    case SIGSEGV:
+        if (rt_bound_error_msg && *rt_bound_error_msg)
+            rt_error(uc, *rt_bound_error_msg);
+        else
+            rt_error(uc, "dereferencing invalid pointer");
+        break;
+    case SIGILL:
+        rt_error(uc, "illegal instruction");
+        break;
+    case SIGABRT:
+        rt_error(uc, "abort() called");
+        break;
+    default:
+        rt_error(uc, "caught signal %d", signum);
+        break;
+    }
+    exit(255);
+}
+
+/* Generate a stack backtrace when a CPU exception occurs. */
+static void set_exception_handler(void)
+{
+    struct sigaction sigact;
+    /* install TCC signal handlers to print debug info on fatal
+       runtime errors */
+    sigact.sa_flags = SA_SIGINFO | SA_RESETHAND;
+    sigact.sa_sigaction = sig_error;
+    sigemptyset(&sigact.sa_mask);
+    sigaction(SIGFPE, &sigact, NULL);
+    sigaction(SIGILL, &sigact, NULL);
+    sigaction(SIGSEGV, &sigact, NULL);
+    sigaction(SIGBUS, &sigact, NULL);
+    sigaction(SIGABRT, &sigact, NULL);
+}
+
+/* ------------------------------------------------------------- */
 #ifdef __i386__
+
 /* fix for glibc 2.1 */
 #ifndef REG_EIP
 #define REG_EIP EIP
@@ -223,8 +400,7 @@ static unsigned long rt_printline(unsigned long wanted_pc)
 #endif
 
 /* return the PC at frame level 'level'. Return non zero if not found */
-static int rt_get_caller_pc(unsigned long *paddr,
-                            ucontext_t *uc, int level)
+static int rt_get_caller_pc(unsigned long *paddr, ucontext_t *uc, int level)
 {
     unsigned long fp;
     int i;
@@ -256,7 +432,10 @@ static int rt_get_caller_pc(unsigned long *paddr,
         return 0;
     }
 }
+
+/* ------------------------------------------------------------- */
 #elif defined(__x86_64__)
+
 /* return the PC at frame level 'level'. Return non zero if not found */
 static int rt_get_caller_pc(unsigned long *paddr,
                             ucontext_t *uc, int level)
@@ -288,236 +467,136 @@ static int rt_get_caller_pc(unsigned long *paddr,
         return 0;
     }
 }
+
+/* ------------------------------------------------------------- */
 #else
+
 #warning add arch specific rt_get_caller_pc()
 static int rt_get_caller_pc(unsigned long *paddr,
                             ucontext_t *uc, int level)
 {
     return -1;
 }
-#endif
 
-/* emit a run time error at position 'pc' */
-void rt_error(ucontext_t *uc, const char *fmt, ...)
-{
-    va_list ap;
-    unsigned long pc;
-    int i;
+#endif /* !__i386__ */
 
-    va_start(ap, fmt);
-    fprintf(stderr, "Runtime error: ");
-    vfprintf(stderr, fmt, ap);
-    fprintf(stderr, "\n");
-    for(i=0;i<num_callers;i++) {
-        if (rt_get_caller_pc(&pc, uc, i) < 0)
-            break;
-        if (i == 0)
-            fprintf(stderr, "at ");
-        else
-            fprintf(stderr, "by ");
-        pc = rt_printline(pc);
-        if (pc == rt_prog_main && pc)
-            break;
-    }
-    exit(255);
-    va_end(ap);
-}
+/* ------------------------------------------------------------- */
+#else /* WIN32 */
 
-/* signal handler for fatal errors */
-static void sig_error(int signum, siginfo_t *siginf, void *puc)
+static long __stdcall cpu_exception_handler(EXCEPTION_POINTERS *ex_info)
 {
-    ucontext_t *uc = puc;
-
-    switch(signum) {
-    case SIGFPE:
-        switch(siginf->si_code) {
-        case FPE_INTDIV:
-        case FPE_FLTDIV:
-            rt_error(uc, "division by zero");
-            break;
-        default:
-            rt_error(uc, "floating point exception");
-            break;
-        }
-        break;
-    case SIGBUS:
-    case SIGSEGV:
-        if (rt_bound_error_msg && *rt_bound_error_msg)
-            rt_error(uc, *rt_bound_error_msg);
-        else
-            rt_error(uc, "dereferencing invalid pointer");
-        break;
-    case SIGILL:
-        rt_error(uc, "illegal instruction");
-        break;
-    case SIGABRT:
-        rt_error(uc, "abort() called");
-        break;
-    default:
-        rt_error(uc, "caught signal %d", signum);
-        break;
-    }
+    CONTEXT *uc = ex_info->ContextRecord;
+/*
+    EXCEPTION_RECORD *er = ex_info->ExceptionRecord;
+    printf("CPU exception: code=%08lx addr=%p\n",
+	er->ExceptionCode, er->ExceptionAddress);
+*/
+    if (rt_bound_error_msg && *rt_bound_error_msg)
+	rt_error(uc, *rt_bound_error_msg);
+    else
+	rt_error(uc, "dereferencing invalid pointer");
     exit(255);
+    //return EXCEPTION_CONTINUE_SEARCH;
 }
 
-#endif /* defined CONFIG_TCC_BACKTRACE */
-
-/********************************************************/
-
-void set_pages_executable(void *ptr, unsigned long length)
+/* Generate a stack backtrace when a CPU exception occurs. */
+static void set_exception_handler(void)
 {
-#ifdef _WIN32
-    unsigned long old_protect;
-    VirtualProtect(ptr, length, PAGE_EXECUTE_READWRITE, &old_protect);
-#else
-    unsigned long start, end;
-    start = (unsigned long)ptr & ~(PAGESIZE - 1);
-    end = (unsigned long)ptr + length;
-    end = (end + PAGESIZE - 1) & ~(PAGESIZE - 1);
-    mprotect((void *)start, end - start, PROT_READ | PROT_WRITE | PROT_EXEC);
-#endif
+    SetUnhandledExceptionFilter(cpu_exception_handler);
 }
 
-/* relocate code. Return -1 on error, required size if ptr is NULL,
-   otherwise copy code into buffer passed by the caller */
-static int tcc_relocate_ex(TCCState *s1, void *ptr)
+/* return the PC at frame level 'level'. Return non zero if not found */
+static int rt_get_caller_pc(unsigned long *paddr,
+			    CONTEXT *uc, int level)
 {
-    Section *s;
-    unsigned long offset, length;
-    uplong mem;
+    unsigned long fp;
     int i;
 
-    if (0 == s1->runtime_added) {
-        s1->runtime_added = 1;
-        s1->nb_errors = 0;
-#ifdef TCC_TARGET_PE
-        pe_output_file(s1, NULL);
-#else
-        tcc_add_runtime(s1);
-        relocate_common_syms();
-        tcc_add_linker_symbols(s1);
-        build_got_entries(s1);
-#endif
-        if (s1->nb_errors)
-            return -1;
-    }
-
-    offset = 0, mem = (uplong)ptr;
-    for(i = 1; i < s1->nb_sections; i++) {
-        s = s1->sections[i];
-        if (0 == (s->sh_flags & SHF_ALLOC))
-            continue;
-        length = s->data_offset;
-        s->sh_addr = mem ? (mem + offset + 15) & ~15 : 0;
-        offset = (offset + length + 15) & ~15;
+    if (level == 0) {
+	*paddr = uc->Eip;
+	return 0;
+    } else {
+	fp = uc->Ebp;
+	for(i=1;i<level;i++) {
+	    /* XXX: check address validity with program info */
+	    if (fp <= 0x1000 || fp >= 0xc0000000)
+		return -1;
+	    fp = ((unsigned long *)fp)[0];
+	}
+	*paddr = ((unsigned long *)fp)[1];
+	return 0;
     }
-    offset += 16;
-
-    /* relocate symbols */
-    relocate_syms(s1, 1);
-    if (s1->nb_errors)
-        return -1;
-
-#ifndef TCC_TARGET_PE
-#ifdef TCC_TARGET_X86_64
-    s1->runtime_plt_and_got_offset = 0;
-    s1->runtime_plt_and_got = (char *)(mem + offset);
-    /* double the size of the buffer for got and plt entries
-       XXX: calculate exact size for them? */
-    offset *= 2;
-#endif
-#endif
+}
 
-    if (0 == mem)
-        return offset;
+#endif /* _WIN32 */
+#endif /* CONFIG_TCC_BACKTRACE */
+/* ------------------------------------------------------------- */
 
-    /* relocate each section */
-    for(i = 1; i < s1->nb_sections; i++) {
-        s = s1->sections[i];
-        if (s->reloc)
-            relocate_section(s1, s);
-    }
+#ifdef CONFIG_TCC_STATIC
 
-    for(i = 1; i < s1->nb_sections; i++) {
-        s = s1->sections[i];
-        if (0 == (s->sh_flags & SHF_ALLOC))
-            continue;
-        length = s->data_offset;
-        // printf("%-12s %08x %04x\n", s->name, s->sh_addr, length);
-        ptr = (void*)(uplong)s->sh_addr;
-        if (NULL == s->data || s->sh_type == SHT_NOBITS)
-            memset(ptr, 0, length);
-        else
-            memcpy(ptr, s->data, length);
-        /* mark executable sections as executable in memory */
-        if (s->sh_flags & SHF_EXECINSTR)
-            set_pages_executable(ptr, length);
-    }
+#define RTLD_LAZY       0x001
+#define RTLD_NOW        0x002
+#define RTLD_GLOBAL     0x100
+#define RTLD_DEFAULT    NULL
 
-#ifndef TCC_TARGET_PE
-#ifdef TCC_TARGET_X86_64
-    set_pages_executable(s1->runtime_plt_and_got,
-                         s1->runtime_plt_and_got_offset);
-#endif
-#endif
-    return 0;
+/* dummy function for profiling */
+void *dlopen(const char *filename, int flag)
+{
+    return NULL;
 }
 
-/* Do all relocations (needed before using tcc_get_symbol())
-   Returns -1 on error. */
-int tcc_relocate(TCCState *s1)
+void dlclose(void *p)
 {
-    int ret = tcc_relocate_ex(s1, NULL);
-    if (-1 == ret)
-        return ret;
-    s1->runtime_mem = tcc_malloc(ret);
-    return tcc_relocate_ex(s1, s1->runtime_mem);
 }
 
-/* launch the compiled program with the given arguments */
-int tcc_run(TCCState *s1, int argc, char **argv)
+const char *dlerror(void)
 {
-    int (*prog_main)(int, char **);
+    return "error";
+}
 
-    if (tcc_relocate(s1) < 0)
-        return -1;
+typedef struct TCCSyms {
+    char *str;
+    void *ptr;
+} TCCSyms;
 
-    prog_main = tcc_get_symbol_err(s1, "main");
+#define TCCSYM(a) { #a, &a, },
 
-    if (s1->do_debug) {
-#ifdef CONFIG_TCC_BACKTRACE
-        struct sigaction sigact;
-        /* install TCC signal handlers to print debug info on fatal
-           runtime errors */
-        sigact.sa_flags = SA_SIGINFO | SA_RESETHAND;
-        sigact.sa_sigaction = sig_error;
-        sigemptyset(&sigact.sa_mask);
-        sigaction(SIGFPE, &sigact, NULL);
-        sigaction(SIGILL, &sigact, NULL);
-        sigaction(SIGSEGV, &sigact, NULL);
-        sigaction(SIGBUS, &sigact, NULL);
-        sigaction(SIGABRT, &sigact, NULL);
-#else
-        error("debug mode not available");
+/* add the symbol you want here if no dynamic linking is done */
+static TCCSyms tcc_syms[] = {
+#if !defined(CONFIG_TCCBOOT)
+    TCCSYM(printf)
+    TCCSYM(fprintf)
+    TCCSYM(fopen)
+    TCCSYM(fclose)
 #endif
+    { NULL, NULL },
+};
+
+void *resolve_sym(TCCState *s1, const char *symbol)
+{
+    TCCSyms *p;
+    p = tcc_syms;
+    while (p->str != NULL) {
+        if (!strcmp(p->str, symbol))
+            return p->ptr;
+        p++;
     }
+    return NULL;
+}
 
-#ifdef CONFIG_TCC_BCHECK
-    if (s1->do_bounds_check) {
-        void (*bound_init)(void);
-        void (*bound_exit)(void);
-        /* set error function */
-        rt_bound_error_msg = tcc_get_symbol_err(s1, "__bound_error_msg");
-        rt_prog_main = (unsigned long)prog_main;
-        /* XXX: use .init section so that it also work in binary ? */
-        bound_init = tcc_get_symbol_err(s1, "__bound_init");
-        bound_exit = tcc_get_symbol_err(s1, "__bound_exit");
-        bound_init();
-        ret = (*prog_main)(argc, argv);
-        bound_exit();
-    } else
-#endif
-    return (*prog_main)(argc, argv);
+#elif defined(_WIN32)
+
+#define dlclose FreeLibrary
+
+#else
+
+#include <dlfcn.h>
+
+void *resolve_sym(TCCState *s1, const char *sym)
+{
+    return dlsym(RTLD_DEFAULT, sym);
 }
 
-/********************************************************/
+#endif /* CONFIG_TCC_STATIC */
+
+/* ------------------------------------------------------------- */
diff --git a/tcctok.h b/tcctok.h
index c214ace..889ba0b 100644
--- a/tcctok.h
+++ b/tcctok.h
@@ -219,9 +219,11 @@
      DEF(TOK___bound_ptr_indir16, "__bound_ptr_indir16")
      DEF(TOK___bound_local_new, "__bound_local_new")
      DEF(TOK___bound_local_delete, "__bound_local_delete")
-#if 0
+#ifdef TCC_TARGET_PE
+#ifndef TCC_TARGET_X86_64
      DEF(TOK_malloc, "malloc")
      DEF(TOK_free, "free")
+#endif
      DEF(TOK_realloc, "realloc")
      DEF(TOK_memalign, "memalign")
      DEF(TOK_calloc, "calloc")
diff --git a/win32/build-tcc.bat b/win32/build-tcc.bat
index b706db4..cb2721a 100644
--- a/win32/build-tcc.bat
+++ b/win32/build-tcc.bat
@@ -31,8 +31,10 @@ copy ..\libtcc.h libtcc\libtcc.h
 :tcc
 %P%gcc -Os -fno-strict-aliasing ../tcc.c -o tcc.exe -s -DTCC_USE_LIBTCC -ltcc -Llibtcc
 
-:libtcc1.a
+:copy_std_includes
 copy ..\include\*.h include
+
+:libtcc1.a
 .\tcc -c lib/crt1.c
 .\tcc -c lib/wincrt1.c
 .\tcc -c lib/dllcrt1.c
@@ -40,5 +42,15 @@ copy ..\include\*.h include
 .\tcc -c ../lib/libtcc1.c
 .\tcc -c lib/chkstk.S
 .\tcc -c ../lib/alloca86%S%.S
-tiny_libmaker lib/libtcc1.a crt1.o wincrt1.o dllcrt1.o dllmain.o chkstk.o libtcc1.o alloca86%S%.o
+
+@set LIBFILES=crt1.o wincrt1.o dllcrt1.o dllmain.o chkstk.o libtcc1.o alloca86%S%.o
+
+@if not _%P%==_ goto makelib
+.\tcc -c ../lib/alloca86-bt%S%.S
+.\tcc -c ../lib/bcheck.c
+
+@set LIBFILES=%LIBFILES% alloca86-bt%S%.o bcheck.o
+
+:makelib
+tiny_libmaker lib/libtcc1.a %LIBFILES%
 del *.o