diff --git a/reactos/include/wine/exception.h b/reactos/include/wine/exception.h
index 594fd338c09..7bfc1c93d55 100644
--- a/reactos/include/wine/exception.h
+++ b/reactos/include/wine/exception.h
@@ -24,6 +24,7 @@
 #include <setjmp.h>
 #include <windef.h>
 #include <excpt.h>
+#include <wine/port.h>
 
 /* The following definitions allow using exceptions in Wine and Winelib code
  *
@@ -75,6 +76,10 @@
 
 #else  /* USE_COMPILER_EXCEPTIONS */
 
+#ifndef __GNUC__
+#define __attribute__(x) /* nothing */
+#endif
+
 #define __TRY \
     do { __WINE_FRAME __f; \
          int __first = 1; \
@@ -91,7 +96,7 @@
              __f.u.filter = (func); \
              __wine_push_frame( &__f.frame ); \
              if (sigsetjmp( __f.jmp, 1 )) { \
-                 const __WINE_FRAME * const __eptr WINE_UNUSED = &__f; \
+                 const __WINE_FRAME * const __eptr __attribute__((unused)) = &__f; \
                  do {
 
 #define __ENDTRY \
@@ -125,10 +130,7 @@ typedef void (CALLBACK *__WINE_FINALLY)(BOOL);
 #define GetExceptionInformation() (__eptr)
 #define GetExceptionCode()        (__eptr->ExceptionRecord->ExceptionCode)
 
-#if defined(AbnormalTermination) && defined(__REACTOS__)
 #undef AbnormalTermination
-#endif 
-
 #define AbnormalTermination()     (!__normal)
 
 typedef struct __tagWINE_FRAME
@@ -175,8 +177,6 @@ static inline EXCEPTION_REGISTRATION_RECORD *__wine_pop_frame( EXCEPTION_REGISTR
 {
 #if defined(__GNUC__) && defined(__i386__)
     __asm__ __volatile__(".byte 0x64\n\tmovl %0,(0)"
-                         //: : "r" (frame->Prev) : "memory" );
-    //return frame->Prev;
                          : : "r" (frame->prev) : "memory" );
     return frame->prev;
 
diff --git a/reactos/tools/winebuild/Makefile b/reactos/tools/winebuild/Makefile
index 6d7cec5f718..2b9596a22f7 100644
--- a/reactos/tools/winebuild/Makefile
+++ b/reactos/tools/winebuild/Makefile
@@ -22,7 +22,7 @@ OBJECTS = \
 
 CLEAN_FILES = *.o $(TARGET)
 
-HOST_CFLAGS = $(HOSTARCH) -D__USE_W32API -I$(PATH_TO_TOP)/include/wine
+HOST_CFLAGS = $(HOSTARCH) -D__USE_W32API -I$(PATH_TO_TOP)/include/wine -I$(PATH_TO_TOP)/include
 
 %.o: %.c
 	$(HOST_CC) $(HOST_CFLAGS) -c $< -o $@
diff --git a/reactos/tools/winebuild/build.h b/reactos/tools/winebuild/build.h
index d6af3a0fbd2..10573db60cf 100644
--- a/reactos/tools/winebuild/build.h
+++ b/reactos/tools/winebuild/build.h
@@ -133,7 +133,7 @@ typedef struct
 #define STACK32OFFSET(reg)  STRUCTOFFSET(STACK32FRAME,reg)
 
   /* Offset of the stack pointer relative to %fs:(0) */
-#define STACKOFFSET (STRUCTOFFSET(TEB,cur_stack))
+#define STACKOFFSET (STRUCTOFFSET(TEB,WOW32Reserved))
 
 
 #define MAX_ORDINALS  65535
@@ -148,6 +148,7 @@ extern void *xmalloc (size_t size);
 extern void *xrealloc (void *ptr, size_t size);
 extern char *xstrdup( const char *str );
 extern char *strupper(char *s);
+extern int strendswith(const char* str, const char* end);
 extern void fatal_error( const char *msg, ... )
    __attribute__ ((__format__ (__printf__, 1, 2)));
 extern void fatal_perror( const char *msg, ... )
@@ -167,16 +168,22 @@ extern void free_dll_spec( DLLSPEC *spec );
 extern const char *make_c_identifier( const char *str );
 extern int get_alignment(int alignBoundary);
 
-extern void add_import_dll( const char *name, int delay );
+extern void add_import_dll( const char *name, const char *filename );
+extern void add_delayed_import( const char *name );
 extern void add_ignore_symbol( const char *name );
 extern void read_undef_symbols( char **argv );
 extern int resolve_imports( DLLSPEC *spec );
-extern int output_imports( FILE *outfile, DLLSPEC *spec );
+extern int output_imports( FILE *outfile, DLLSPEC *spec, int *nb_delayed );
 extern int load_res32_file( const char *name, DLLSPEC *spec );
 extern void output_resources( FILE *outfile, DLLSPEC *spec );
 extern void load_res16_file( const char *name, DLLSPEC *spec );
-extern int output_res16_data( FILE *outfile, DLLSPEC *spec );
-extern int output_res16_directory( unsigned char *buffer, DLLSPEC *spec );
+extern unsigned int get_res16_data_size( DLLSPEC *spec, unsigned int res_offset,
+                                         unsigned int alignment );
+extern unsigned int output_res16_data( unsigned char **ret_buf, DLLSPEC *spec,
+                                       unsigned int res_offset, unsigned int alignment );
+extern unsigned int get_res16_directory_size( DLLSPEC *spec );
+extern unsigned int output_res16_directory( unsigned char **ret_buf, DLLSPEC *spec,
+                                            unsigned int res_offset, unsigned int alignment );
 extern void output_dll_init( FILE *outfile, const char *constructor, const char *destructor );
 
 extern void BuildRelays16( FILE *outfile );
diff --git a/reactos/tools/winebuild/import.c b/reactos/tools/winebuild/import.c
index 0a6aff1b3d1..e6f97e144db 100644
--- a/reactos/tools/winebuild/import.c
+++ b/reactos/tools/winebuild/import.c
@@ -26,16 +26,23 @@
 #include <stdio.h>
 #include <string.h>
 #include <stdarg.h>
+#ifdef HAVE_SYS_STAT_H
+# include <sys/stat.h>
+#endif
 #ifdef HAVE_UNISTD_H
 # include <unistd.h>
 #endif
 
 #include "winglue.h"
+#include "wine/exception.h"
 #include "build.h"
 
 struct import
 {
-    DLLSPEC     *spec;         /* description of the imported dll */
+    DLLSPEC     *spec;        /* description of the imported dll */
+    char        *full_name;   /* full name of the input file */
+    dev_t        dev;         /* device/inode of the input file */
+    ino_t        ino;
     int          delay;       /* delay or not dll loading ? */
     ORDDEF     **exports;     /* functions exported from this dll */
     int          nb_exports;  /* number of exported functions */
@@ -58,6 +65,8 @@ static int nb_imports = 0;      /* number of imported dlls (delayed or not) */
 static int nb_delayed = 0;      /* number of delayed dlls */
 static int total_imports = 0;   /* total number of imported functions */
 static int total_delayed = 0;   /* total number of imported functions in delayed DLLs */
+static char **delayed_imports;  /* names of delayed import dlls */
+static int nb_delayed_imports;  /* size of the delayed_imports array */
 
 /* list of symbols that are ignored by default */
 static const char * const default_ignored_symbols[] =
@@ -187,6 +196,7 @@ static void free_imports( struct import *imp )
     free( imp->exports );
     free( imp->imports );
     free_dll_spec( imp->spec );
+    free( imp->full_name );
     free( imp );
 }
 
@@ -196,16 +206,28 @@ static void remove_ld_tmp_file(void)
     if (ld_tmp_file) unlink( ld_tmp_file );
 }
 
+/* check whether a given dll is imported in delayed mode */
+static int is_delayed_import( const char *name )
+{
+    int i;
+
+    for (i = 0; i < nb_delayed_imports; i++)
+    {
+        if (!strcmp( delayed_imports[i], name )) return 1;
+    }
+    return 0;
+}
+
 /* check whether a given dll has already been imported */
-static int is_already_imported( const char *name )
+static struct import *is_already_imported( const char *name )
 {
     int i;
 
     for (i = 0; i < nb_imports; i++)
     {
-        if (!strcmp( dll_imports[i]->spec->file_name, name )) return 1;
+        if (!strcmp( dll_imports[i]->spec->file_name, name )) return dll_imports[i];
     }
-    return 0;
+    return NULL;
 }
 
 /* open the .so library for a given dll in a specified path */
@@ -227,8 +249,8 @@ static char *try_library_path( const char *path, const char *name )
     return NULL;
 }
 
-/* open the .so library for a given dll */
-static char *open_library( const char *name )
+/* find the .def import library for a given dll */
+static char *find_library( const char *name )
 {
     char *fullname;
     int i;
@@ -242,24 +264,36 @@ static char *open_library( const char *name )
 }
 
 /* read in the list of exported symbols of an import library */
-static int read_import_lib( const char *name, struct import *imp )
+static int read_import_lib( struct import *imp )
 {
     FILE *f;
-    char *fullname;
     int i, ret;
+    struct stat stat;
+    struct import *prev_imp;
     DLLSPEC *spec = imp->spec;
 
-    imp->exports    = NULL;
-    imp->nb_exports = 0;
-
-    fullname = open_library( name );
-    f = open_input_file( NULL, fullname );
-    free( fullname );
-
+    f = open_input_file( NULL, imp->full_name );
+    fstat( fileno(f), &stat );
+    imp->dev = stat.st_dev;
+    imp->ino = stat.st_ino;
     ret = parse_def_file( f, spec );
     close_input_file( f );
     if (!ret) return 0;
-    if (is_already_imported( spec->file_name )) return 0;
+
+    /* check if we already imported that library from a different file */
+    if ((prev_imp = is_already_imported( spec->file_name )))
+    {
+        if (prev_imp->dev != imp->dev || prev_imp->ino != imp->ino)
+            fatal_error( "%s and %s have the same export name '%s'\n",
+                         prev_imp->full_name, imp->full_name, spec->file_name );
+        return 0;  /* the same file was already loaded, ignore this one */
+    }
+
+    if (is_delayed_import( spec->file_name ))
+    {
+        imp->delay = 1;
+        nb_delayed++;
+    }
 
     imp->exports = xmalloc( spec->nb_entry_points * sizeof(*imp->exports) );
 
@@ -277,32 +311,47 @@ static int read_import_lib( const char *name, struct import *imp )
     return 1;
 }
 
-/* add a dll to the list of imports */
-void add_import_dll( const char *name, int delay )
+/* build the dll exported name from the import lib name or path */
+static char *get_dll_name( const char *name, const char *filename )
 {
-    struct import *imp;
-    char *fullname;
+    char *ret;
 
-    fullname = xmalloc( strlen(name) + 5 );
-    strcpy( fullname, name );
-    if (!strchr( fullname, '.' )) strcat( fullname, ".dll" );
-
-    /* check if we already imported it */
-    if (is_already_imported( fullname ))
+    if (filename)
     {
-        free( fullname );
-        return;
+        const char *basename = strrchr( filename, '/' );
+        if (!basename) basename = filename;
+        else basename++;
+        if (!strncmp( basename, "lib", 3 )) basename += 3;
+        ret = xmalloc( strlen(basename) + 5 );
+        strcpy( ret, basename );
+        if (strendswith( ret, ".def" )) ret[strlen(ret)-4] = 0;
     }
+    else
+    {
+        ret = xmalloc( strlen(name) + 5 );
+        strcpy( ret, name );
+    }
+    if (!strchr( ret, '.' )) strcat( ret, ".dll" );
+    return ret;
+}
+
+/* add a dll to the list of imports */
+void add_import_dll( const char *name, const char *filename )
+{
+    struct import *imp = xmalloc( sizeof(*imp) );
 
-    imp = xmalloc( sizeof(*imp) );
     imp->spec            = alloc_dll_spec();
-    imp->spec->file_name = fullname;
-    imp->delay           = delay;
+    imp->spec->file_name = get_dll_name( name, filename );
+    imp->delay           = 0;
     imp->imports         = NULL;
     imp->nb_imports      = 0;
-    if (delay) nb_delayed++;
+    imp->exports         = NULL;
+    imp->nb_exports      = 0;
 
-    if (read_import_lib( name, imp ))
+    if (filename) imp->full_name = xstrdup( filename );
+    else imp->full_name = find_library( name );
+
+    if (read_import_lib( imp ))
     {
         dll_imports = xrealloc( dll_imports, (nb_imports+1) * sizeof(*dll_imports) );
         dll_imports[nb_imports++] = imp;
@@ -314,6 +363,21 @@ void add_import_dll( const char *name, int delay )
     }
 }
 
+/* add a library to the list of delayed imports */
+void add_delayed_import( const char *name )
+{
+    struct import *imp;
+    char *fullname = get_dll_name( name, NULL );
+
+    delayed_imports = xrealloc( delayed_imports, (nb_delayed_imports+1) * sizeof(*delayed_imports) );
+    delayed_imports[nb_delayed_imports++] = fullname;
+    if ((imp = is_already_imported( fullname )) && !imp->delay)
+    {
+        imp->delay = 1;
+        nb_delayed++;
+    }
+}
+
 /* remove an imported dll, based on its index in the dll_imports array */
 static void remove_import_dll( int index )
 {
@@ -438,7 +502,7 @@ static int add_extra_symbol( const char **extras, int *count, const char *name,
 static void add_extra_undef_symbols( const DLLSPEC *spec )
 {
     const char *extras[10];
-    int i, count = 0, nb_stubs = 0, nb_regs = 0;
+    int i, count = 0, nb_stubs = 0;
     int kernel_imports = 0, ntdll_imports = 0;
 
     sort_symbols( undef_symbols, nb_undef_symbols );
@@ -447,7 +511,6 @@ static void add_extra_undef_symbols( const DLLSPEC *spec )
     {
         ORDDEF *odp = &spec->entry_points[i];
         if (odp->type == TYPE_STUB) nb_stubs++;
-        if (odp->flags & FLAG_REGISTER) nb_regs++;
     }
 
     /* add symbols that will be contained in the spec file itself */
@@ -467,16 +530,16 @@ static void add_extra_undef_symbols( const DLLSPEC *spec )
     if (nb_delayed)
     {
         kernel_imports += add_extra_symbol( extras, &count, "LoadLibraryA", spec );
+        kernel_imports += add_extra_symbol( extras, &count, "FreeLibrary", spec );
         kernel_imports += add_extra_symbol( extras, &count, "GetProcAddress", spec );
+        kernel_imports += add_extra_symbol( extras, &count, "RaiseException", spec );
     }
-    if (nb_regs)
-        ntdll_imports += add_extra_symbol( extras, &count, "__wine_call_from_32_regs", spec );
-    if (nb_delayed || nb_stubs)
+    if (nb_stubs)
         ntdll_imports += add_extra_symbol( extras, &count, "RtlRaiseException", spec );
 
     /* make sure we import the dlls that contain these functions */
-    if (kernel_imports) add_import_dll( "kernel32", 0 );
-    if (ntdll_imports) add_import_dll( "ntdll", 0 );
+    if (kernel_imports) add_import_dll( "kernel32", NULL );
+    if (ntdll_imports) add_import_dll( "ntdll", NULL );
 
     if (count)
     {
@@ -777,10 +840,10 @@ static int output_delayed_imports( FILE *outfile, const DLLSPEC *spec )
 
     if (!nb_delayed) goto done;
 
+    fprintf( outfile, "static void *__wine_delay_imp_hmod[%d];\n", nb_delayed );
     for (i = 0; i < nb_imports; i++)
     {
         if (!dll_imports[i]->delay) continue;
-        fprintf( outfile, "static void *__wine_delay_imp_%d_hmod;\n", i);
         for (j = 0; j < dll_imports[i]->nb_imports; j++)
         {
             ORDDEF *odp = dll_imports[i]->imports[j];
@@ -807,7 +870,7 @@ static int output_delayed_imports( FILE *outfile, const DLLSPEC *spec )
     for (i = j = 0; i < nb_imports; i++)
     {
         if (!dll_imports[i]->delay) continue;
-        fprintf( outfile, "    { 0, \"%s\", &__wine_delay_imp_%d_hmod, &delay_imports.IAT[%d], &delay_imports.INT[%d], 0, 0, 0 },\n",
+        fprintf( outfile, "    { 0, \"%s\", &__wine_delay_imp_hmod[%d], &delay_imports.IAT[%d], &delay_imports.INT[%d], 0, 0, 0 },\n",
                  dll_imports[i]->spec->file_name, i, j, j );
         j += dll_imports[i]->nb_imports;
     }
@@ -839,21 +902,7 @@ static int output_delayed_imports( FILE *outfile, const DLLSPEC *spec )
     }
     fprintf( outfile, "  }\n};\n\n" );
 
-    /* check if there's some stub defined. if so, exception struct
-     *  is already defined, so don't emit it twice
-     */
-    for (i = 0; i < spec->nb_entry_points; i++) if (spec->entry_points[i].type == TYPE_STUB) break;
-
-    if (i == spec->nb_entry_points) {
-       fprintf( outfile, "struct exc_record {\n" );
-       fprintf( outfile, "  unsigned int code, flags;\n" );
-       fprintf( outfile, "  void *rec, *addr;\n" );
-       fprintf( outfile, "  unsigned int params;\n" );
-       fprintf( outfile, "  const void *info[15];\n" );
-       fprintf( outfile, "};\n\n" );
-       fprintf( outfile, "extern void __stdcall RtlRaiseException( struct exc_record * );\n" );
-    }
-
+    fprintf( outfile, "extern void __stdcall RaiseException(unsigned int, unsigned int, unsigned int, const void *args[]);\n" );
     fprintf( outfile, "extern void * __stdcall LoadLibraryA(const char*);\n");
     fprintf( outfile, "extern void * __stdcall GetProcAddress(void *, const char*);\n");
     fprintf( outfile, "\n" );
@@ -871,20 +920,12 @@ static int output_delayed_imports( FILE *outfile, const DLLSPEC *spec )
     fprintf( outfile, "    /* patch IAT with final value */\n" );
     fprintf( outfile, "    return *pIAT = fn;\n" );
     fprintf( outfile, "  else {\n");
-    fprintf( outfile, "    struct exc_record rec;\n" );
-    fprintf( outfile, "    rec.code    = 0x80000100;\n" );
-    fprintf( outfile, "    rec.flags   = 1;\n" );
-    fprintf( outfile, "    rec.rec     = 0;\n" );
-    fprintf( outfile, "    rec.params  = 2;\n" );
-    fprintf( outfile, "    rec.info[0] = imd->szName;\n" );
-    fprintf( outfile, "    rec.info[1] = *pINT;\n" );
-    fprintf( outfile, "#ifdef __GNUC__\n" );
-    fprintf( outfile, "    rec.addr = __builtin_return_address(1);\n" );
-    fprintf( outfile, "#else\n" );
-    fprintf( outfile, "    rec.addr = 0;\n" );
-    fprintf( outfile, "#endif\n" );
-    fprintf( outfile, "    for (;;) RtlRaiseException( &rec );\n" );
-    fprintf( outfile, "    return 0; /* shouldn't go here */\n" );
+    fprintf( outfile, "    const void *args[2];\n" );
+    fprintf( outfile, "    args[0] = imd->szName;\n" );
+    fprintf( outfile, "    args[1] = *pINT;\n" );
+    fprintf( outfile, "    RaiseException( 0x%08x, %d, 2, args );\n",
+             EXCEPTION_WINE_STUB, EXCEPTION_NONCONTINUABLE );
+    fprintf( outfile, "    return 0;\n" );
     fprintf( outfile, "  }\n}\n\n" );
 
     fprintf( outfile, "#ifndef __GNUC__\n" );
@@ -1100,8 +1141,8 @@ static int output_delayed_imports( FILE *outfile, const DLLSPEC *spec )
 /* output the import and delayed import tables of a Win32 module
  * returns number of DLLs exported in 'immediate' mode
  */
-int output_imports( FILE *outfile, DLLSPEC *spec )
+int output_imports( FILE *outfile, DLLSPEC *spec, int *nb_delayed )
 {
-   output_delayed_imports( outfile, spec );
-   return output_immediate_imports( outfile );
+    *nb_delayed = output_delayed_imports( outfile, spec );
+    return output_immediate_imports( outfile );
 }
diff --git a/reactos/tools/winebuild/main.c b/reactos/tools/winebuild/main.c
index 8148ba71677..3f98a88a46c 100644
--- a/reactos/tools/winebuild/main.c
+++ b/reactos/tools/winebuild/main.c
@@ -283,7 +283,7 @@ static char **parse_options( int argc, char **argv, DLLSPEC *spec )
             spec->dll_name = xstrdup( optarg );
             break;
         case 'd':
-            add_import_dll( optarg, 1 );
+            add_delayed_import( optarg );
             break;
         case 'e':
             spec->init_func = xstrdup( optarg );
@@ -312,7 +312,7 @@ static char **parse_options( int argc, char **argv, DLLSPEC *spec )
             kill_at = 1;
             break;
         case 'l':
-            add_import_dll( optarg, 0 );
+            add_import_dll( optarg, NULL );
             break;
         case 'o':
             if (unlink( optarg ) == -1 && errno != ENOENT)
@@ -408,6 +408,21 @@ static void load_resources( char *argv[], DLLSPEC *spec )
     }
 }
 
+/* add input files that look like import libs to the import list */
+static void load_import_libs( char *argv[] )
+{
+    char **ptr, **last;
+
+    for (ptr = last = argv; *ptr; ptr++)
+    {
+        if (strendswith( *ptr, ".def" ))
+            add_import_dll( NULL, *ptr );
+        else
+            *last++ = *ptr; /* not an import dll, keep it in the list */
+    }
+    *last = NULL;
+}
+
 static int parse_input_file( DLLSPEC *spec )
 {
     FILE *input_file = open_input_file( NULL, spec_file_name );
@@ -444,6 +459,7 @@ int main(int argc, char **argv)
     case MODE_DLL:
         spec->characteristics |= IMAGE_FILE_DLL;
         load_resources( argv, spec );
+        load_import_libs( argv );
         if (!spec_file_name) fatal_error( "missing .spec file\n" );
         if (!parse_input_file( spec )) break;
         switch (spec->type)
@@ -462,6 +478,7 @@ int main(int argc, char **argv)
         if (spec->type == SPEC_WIN16) fatal_error( "Cannot build 16-bit exe files\n" );
 	if (!spec->file_name) fatal_error( "executable must be named via the -F option\n" );
         load_resources( argv, spec );
+        load_import_libs( argv );
         if (spec_file_name && !parse_input_file( spec )) break;
         read_undef_symbols( argv );
         BuildSpec32File( output_file, spec );
diff --git a/reactos/tools/winebuild/parser.c b/reactos/tools/winebuild/parser.c
index a998777d873..424a0ef499b 100644
--- a/reactos/tools/winebuild/parser.c
+++ b/reactos/tools/winebuild/parser.c
@@ -511,6 +511,12 @@ static int parse_spec_ordinal( int ordinal, DLLSPEC *spec )
         odp->ordinal = ordinal;
     }
 
+    if (spec->type == SPEC_WIN32 && odp->flags & FLAG_REGISTER)
+    {
+        error( "-register flag not supported for Win32 entry points\n" );
+        goto error;
+    }
+
     if (odp->type == TYPE_STDCALL && !(odp->flags & FLAG_PRIVATE))
     {
         if (!strcmp( odp->name, "DllRegisterServer" ) ||
diff --git a/reactos/tools/winebuild/relay.c b/reactos/tools/winebuild/relay.c
index 6548714b43d..abac75ae4d9 100644
--- a/reactos/tools/winebuild/relay.c
+++ b/reactos/tools/winebuild/relay.c
@@ -27,7 +27,7 @@
 #include <ctype.h>
 
 #include "thread.h"
-#include "stackframe.h"
+#include "wine/winbase16.h"
 
 #include "build.h"
 
diff --git a/reactos/tools/winebuild/res16.c b/reactos/tools/winebuild/res16.c
index a5fd5849132..87a16029687 100644
--- a/reactos/tools/winebuild/res16.c
+++ b/reactos/tools/winebuild/res16.c
@@ -20,6 +20,7 @@
 
 #include "config.h"
 
+#include <assert.h>
 #include <ctype.h>
 #include <stdlib.h>
 #include <string.h>
@@ -39,9 +40,6 @@
 #include "winglue.h"
 #include "build.h"
 
-#define ALIGNMENT 2 /* alignment for resource data */
-#define ALIGN_MASK ((1 << ALIGNMENT) - 1)
-
 /* Unicode string or integer id */
 struct string_id
 {
@@ -265,48 +263,115 @@ static void output_string( unsigned char **buffer, const char *str )
     while (len--) put_byte( buffer, *str++ );
 }
 
-/* output the resource data */
-int output_res16_data( FILE *outfile, DLLSPEC *spec )
+/* get the resource data total size */
+unsigned int get_res16_data_size( DLLSPEC *spec, unsigned int res_offset, unsigned int alignment )
 {
     const struct resource *res;
-    unsigned char *buffer, *p;
-    unsigned int i;
-    int total;
+    unsigned int i, total;
+    unsigned int align_mask = (1 << alignment) - 1;
 
     if (!spec->nb_resources) return 0;
 
-    for (i = total = 0, res = spec->resources; i < spec->nb_resources; i++, res++)
-        total += (res->data_size + ALIGN_MASK) & ~ALIGN_MASK;
+    /* add padding at the beginning if needed so that resources are properly aligned */
+    total = ((res_offset + align_mask) & ~align_mask) - res_offset;
 
-    buffer = p = xmalloc( total );
     for (i = 0, res = spec->resources; i < spec->nb_resources; i++, res++)
-    {
-        memcpy( p, res->data, res->data_size );
-        p += res->data_size;
-        while ((int)p & ALIGN_MASK) *p++ = 0;
-    }
-    dump_bytes( outfile, buffer, total, "resource_data", 1 );
-    free( buffer );
+        total += (res->data_size + align_mask) & ~align_mask;
+
     return total;
 }
 
-/* output the resource definitions */
-int output_res16_directory( unsigned char *buffer, DLLSPEC *spec )
+/* output the resource data */
+unsigned int output_res16_data( unsigned char **ret_buf, DLLSPEC *spec,
+                                unsigned int res_offset, unsigned int alignment )
 {
-    int offset, res_offset = 0;
-    unsigned int i, j;
+    const struct resource *res;
+    unsigned char *p;
+    unsigned int i, total, padding;
+    unsigned int align_mask = (1 << alignment) - 1;
+
+    if (!spec->nb_resources) return 0;
+
+    /* add padding at the beginning if needed so that resources are properly aligned */
+    padding = ((res_offset + align_mask) & ~align_mask) - res_offset;
+
+    for (i = total = 0, res = spec->resources; i < spec->nb_resources; i++, res++)
+        total += (res->data_size + align_mask) & ~align_mask;
+
+    *ret_buf = p = xmalloc( total + padding );
+    memset( p, 0, padding );
+    p += padding;
+    for (i = 0, res = spec->resources; i < spec->nb_resources; i++, res++)
+    {
+        unsigned int size = (res->data_size + align_mask) & ~align_mask;
+        memcpy( p, res->data, res->data_size );
+        memset( p + res->data_size, 0, size - res->data_size );
+        p += size;
+    }
+    return total;
+}
+
+/* get the resource definitions total size */
+unsigned int get_res16_directory_size( DLLSPEC *spec )
+{
+    unsigned int i, j, total_size;
     struct res_tree *tree;
     const struct res_type *type;
     const struct resource *res;
-    unsigned char *start = buffer;
 
     tree = build_resource_tree( spec );
 
+    total_size = 4;  /* alignment + terminator */
+    total_size += tree->nb_types * 8;  /* typeinfo structures */
+    total_size += spec->nb_resources * 12;  /* nameinfo structures */
+
+    for (i = 0, type = tree->types; i < tree->nb_types; i++, type++)
+    {
+        if (type->type->str) total_size += strlen(type->type->str) + 1;
+        for (j = 0, res = type->res; j < type->nb_names; j++, res++)
+            if (res->name.str) total_size += strlen(res->name.str) + 1;
+    }
+    total_size++;  /* final terminator */
+    if (total_size & 1) total_size++;
+    return total_size;
+}
+
+/* output the resource definitions */
+unsigned int output_res16_directory( unsigned char **ret_buf, DLLSPEC *spec,
+                                     unsigned int res_offset, unsigned int alignment )
+{
+    int offset;
+    unsigned int i, j, total_size;
+    unsigned int align_mask = (1 << alignment) - 1;
+    struct res_tree *tree;
+    const struct res_type *type;
+    const struct resource *res;
+    unsigned char *buffer;
+
+    tree = build_resource_tree( spec );
+
+    /* make sure data offset is properly aligned */
+    res_offset = (res_offset + align_mask) & ~align_mask;
+
+    /* first compute total size */
+
     offset = 4;  /* alignment + terminator */
     offset += tree->nb_types * 8;  /* typeinfo structures */
     offset += spec->nb_resources * 12;  /* nameinfo structures */
 
-    put_word( &buffer, ALIGNMENT );
+    total_size = offset;
+
+    for (i = 0, type = tree->types; i < tree->nb_types; i++, type++)
+    {
+        if (type->type->str) total_size += strlen(type->type->str) + 1;
+        for (j = 0, res = type->res; j < type->nb_names; j++, res++)
+            if (res->name.str) total_size += strlen(res->name.str) + 1;
+    }
+    total_size++;  /* final terminator */
+    if (total_size & 1) total_size++;
+    *ret_buf = buffer = xmalloc( total_size );
+
+    put_word( &buffer, alignment );
 
     /* type and name structures */
 
@@ -326,8 +391,8 @@ int output_res16_directory( unsigned char *buffer, DLLSPEC *spec )
 
         for (j = 0, res = type->res; j < type->nb_names; j++, res++)
         {
-            put_word( &buffer, res_offset >> ALIGNMENT );
-            put_word( &buffer, (res->data_size + ALIGN_MASK) >> ALIGNMENT );
+            put_word( &buffer, res_offset >> alignment );
+            put_word( &buffer, (res->data_size + align_mask) >> alignment );
             put_word( &buffer, res->memopt );
             if (res->name.str)
             {
@@ -338,7 +403,7 @@ int output_res16_directory( unsigned char *buffer, DLLSPEC *spec )
                 put_word( &buffer, res->name.id | 0x8000 );
             put_word( &buffer, 0 );
             put_word( &buffer, 0 );
-            res_offset += (res->data_size + ALIGN_MASK) & ~ALIGN_MASK;
+            res_offset += (res->data_size + align_mask) & ~align_mask;
         }
     }
     put_word( &buffer, 0 );  /* terminator */
@@ -354,8 +419,9 @@ int output_res16_directory( unsigned char *buffer, DLLSPEC *spec )
         }
     }
     put_byte( &buffer, 0 );  /* names terminator */
-    if ((buffer - start) & 1) put_byte( &buffer, 0 );  /* align on word boundary */
+    if ((buffer - *ret_buf) & 1) put_byte( &buffer, 0 );  /* align on word boundary */
+    assert( buffer - *ret_buf == total_size );
 
     free_resource_tree( tree );
-    return buffer - start;
+    return total_size;
 }
diff --git a/reactos/tools/winebuild/spec16.c b/reactos/tools/winebuild/spec16.c
index aae4d8f5b97..d8871ea0050 100644
--- a/reactos/tools/winebuild/spec16.c
+++ b/reactos/tools/winebuild/spec16.c
@@ -28,9 +28,7 @@
 #include <ctype.h>
 
 #include "wine/exception.h"
-#include "stackframe.h"
-#include "builtin16.h"
-#include "module.h"
+#include "wine/winbase16.h"
 
 #include "build.h"
 
@@ -76,230 +74,95 @@ static void output_file_header( FILE *outfile )
 
 
 /*******************************************************************
- *         StoreVariableCode
- *
- * Store a list of ints into a byte array.
+ *         output_entry_table
  */
-static int StoreVariableCode( unsigned char *buffer, int size, ORDDEF *odp )
+static int output_entry_table( unsigned char **ret_buff, DLLSPEC *spec )
 {
-    int i;
+    int i, prev = 0, prev_sel = -1;
+    unsigned char *pstr, *buffer;
+    unsigned char *bundle = NULL;
 
-    switch(size)
-    {
-    case 1:
-        for (i = 0; i < odp->u.var.n_values; i++)
-            buffer[i] = odp->u.var.values[i];
-        break;
-    case 2:
-        for (i = 0; i < odp->u.var.n_values; i++)
-            ((unsigned short *)buffer)[i] = odp->u.var.values[i];
-        break;
-    case 4:
-        for (i = 0; i < odp->u.var.n_values; i++)
-            ((unsigned int *)buffer)[i] = odp->u.var.values[i];
-        break;
-    }
-    return odp->u.var.n_values * size;
-}
+    buffer = xmalloc( spec->limit * 5 );  /* we use at most 5 bytes per entry-point */
+    pstr = buffer;
 
-
-/*******************************************************************
- *         BuildModule16
- *
- * Build the in-memory representation of a 16-bit NE module, and dump it
- * as a byte stream into the assembly code.
- */
-static int BuildModule16( FILE *outfile, int max_code_offset,
-                          int max_data_offset, DLLSPEC *spec )
-{
-    int i;
-    char *buffer;
-    NE_MODULE *pModule;
-    SEGTABLEENTRY *pSegment;
-    OFSTRUCT *pFileInfo;
-    BYTE *pstr;
-    ET_BUNDLE *bundle = 0;
-    ET_ENTRY entry;
-
-    /*   Module layout:
-     * NE_MODULE       Module
-     * OFSTRUCT        File information
-     * SEGTABLEENTRY   Segment 1 (code)
-     * SEGTABLEENTRY   Segment 2 (data)
-     * WORD[2]         Resource table (empty)
-     * BYTE[2]         Imported names (empty)
-     * BYTE[n]         Resident names table
-     * BYTE[n]         Entry table
-     */
-
-    buffer = xmalloc( 0x10000 );
-    memset( buffer, 0, 0x10000 );
-
-    pModule = (NE_MODULE *)buffer;
-    pModule->magic = IMAGE_OS2_SIGNATURE;
-    pModule->count = 1;
-    pModule->next = 0;
-    pModule->flags = NE_FFLAGS_SINGLEDATA | NE_FFLAGS_BUILTIN | NE_FFLAGS_LIBMODULE;
-    pModule->dgroup = 2;
-    pModule->heap_size = spec->heap_size;
-    pModule->stack_size = 0;
-    pModule->ip = 0;
-    pModule->cs = 0;
-    pModule->sp = 0;
-    pModule->ss = 0;
-    pModule->seg_count = 2;
-    pModule->modref_count = 0;
-    pModule->nrname_size = 0;
-    pModule->modref_table = 0;
-    pModule->nrname_fpos = 0;
-    pModule->moveable_entries = 0;
-    pModule->alignment = 0;
-    pModule->truetype = 0;
-    pModule->os_flags = NE_OSFLAGS_WINDOWS;
-    pModule->misc_flags = 0;
-    pModule->dlls_to_init  = 0;
-    pModule->nrname_handle = 0;
-    pModule->min_swap_area = 0;
-    pModule->expected_version = 0;
-    pModule->module32 = 0;
-    pModule->self = 0;
-    pModule->self_loading_sel = 0;
-
-      /* File information */
-
-    pFileInfo = (OFSTRUCT *)(pModule + 1);
-    pModule->fileinfo = (int)pFileInfo - (int)pModule;
-    memset( pFileInfo, 0, sizeof(*pFileInfo) - sizeof(pFileInfo->szPathName) );
-    pFileInfo->cBytes = sizeof(*pFileInfo) - sizeof(pFileInfo->szPathName)
-                        + strlen(spec->file_name);
-    strcpy( pFileInfo->szPathName, spec->file_name );
-    pstr = (char *)pFileInfo + pFileInfo->cBytes + 1;
-
-      /* Segment table */
-
-    pstr = (char *)(((long)pstr + 3) & ~3);
-    pSegment = (SEGTABLEENTRY *)pstr;
-    pModule->seg_table = (int)pSegment - (int)pModule;
-    pSegment->filepos = 0;
-    pSegment->size = max_code_offset;
-    pSegment->flags = 0;
-    pSegment->minsize = max_code_offset;
-    pSegment->hSeg = 0;
-    pSegment++;
-
-    pModule->dgroup_entry = (int)pSegment - (int)pModule;
-    pSegment->filepos = 0;
-    pSegment->size = max_data_offset;
-    pSegment->flags = NE_SEGFLAGS_DATA;
-    pSegment->minsize = max_data_offset;
-    pSegment->hSeg = 0;
-    pSegment++;
-
-      /* Resource table */
-
-    pstr = (char *)pSegment;
-    pstr = (char *)(((long)pstr + 3) & ~3);
-    pModule->res_table = (int)pstr - (int)pModule;
-    pstr += output_res16_directory( pstr, spec );
-
-      /* Imported names table */
-
-    pstr = (char *)(((long)pstr + 3) & ~3);
-    pModule->import_table = (int)pstr - (int)pModule;
-    *pstr++ = 0;
-    *pstr++ = 0;
-
-      /* Resident names table */
-
-    pstr = (char *)(((long)pstr + 3) & ~3);
-    pModule->name_table = (int)pstr - (int)pModule;
-    /* First entry is module name */
-    *pstr = strlen( spec->dll_name );
-    strcpy( pstr + 1, spec->dll_name );
-    strupper( pstr + 1 );
-    pstr += *pstr + 1;
-    *pstr++ = 0;
-    *pstr++ = 0;
-    /* Store all ordinals */
-    for (i = 1; i <= spec->limit; i++)
-    {
-        ORDDEF *odp = spec->ordinals[i];
-        WORD ord = i;
-        if (!odp || !odp->name[0]) continue;
-        *pstr = strlen( odp->name );
-        strcpy( pstr + 1, odp->name );
-        strupper( pstr + 1 );
-        pstr += *pstr + 1;
-        memcpy( pstr, &ord, sizeof(WORD) );
-        pstr += sizeof(WORD);
-    }
-    *pstr++ = 0;
-
-      /* Entry table */
-
-    pstr = (char *)(((long)pstr + 3) & ~3);
-    pModule->entry_table = (int)pstr - (int)pModule;
     for (i = 1; i <= spec->limit; i++)
     {
         int selector = 0;
+        WORD offset;
         ORDDEF *odp = spec->ordinals[i];
         if (!odp) continue;
 
-	switch (odp->type)
-	{
+        switch (odp->type)
+        {
         case TYPE_CDECL:
         case TYPE_PASCAL:
         case TYPE_VARARGS:
         case TYPE_STUB:
             selector = 1;  /* Code selector */
             break;
-
         case TYPE_VARIABLE:
             selector = 2;  /* Data selector */
             break;
-
         case TYPE_ABS:
             selector = 0xfe;  /* Constant selector */
             break;
-
         default:
-            selector = 0;  /* Invalid selector */
-            break;
+            continue;
         }
 
-        if ( !selector )
-           continue;
-
-        if ( bundle && bundle->last+1 == i )
-            bundle->last++;
-        else
+        if (!bundle || prev + 1 != i || prev_sel != selector || *bundle == 255)
         {
-            pstr = (char *)(((long)pstr + 1) & ~1);
-            if ( bundle )
-                bundle->next = (char *)pstr - (char *)pModule;
+            /* need to start a new bundle */
 
-            bundle = (ET_BUNDLE *)pstr;
-            bundle->first = i-1;
-            bundle->last = i;
-            bundle->next = 0;
-            pstr += sizeof(ET_BUNDLE);
+            if (prev + 1 != i)
+            {
+                int skip = i - (prev + 1);
+                while (skip > 255)
+                {
+                    *pstr++ = 255;
+                    *pstr++ = 0;
+                    skip -= 255;
+                }
+                *pstr++ = skip;
+                *pstr++ = 0;
+            }
+
+            bundle = pstr;
+            *pstr++ = 0;
+            *pstr++ = selector;
+            prev_sel = selector;
         }
-
-	/* FIXME: is this really correct ?? */
-	entry.type = 0xff;  /* movable */
-	entry.flags = 3; /* exported & public data */
-	entry.segnum = selector;
-	entry.offs = odp->offset;
-        memcpy( pstr, &entry, sizeof(ET_ENTRY) );
-	pstr += sizeof(ET_ENTRY);
+        /* output the entry */
+        *pstr++ = 3;  /* flags: exported & public data */
+        offset = odp->offset;
+        memcpy( pstr, &offset, sizeof(WORD) );
+        pstr += sizeof(WORD);
+        (*bundle)++;  /* increment bundle entry count */
+        prev = i;
     }
     *pstr++ = 0;
+    if ((pstr - buffer) & 1) *pstr++ = 0;
+    *ret_buff = xrealloc( buffer, pstr - buffer );
+    return pstr - buffer;
+}
 
-      /* Dump the module content */
 
-    pstr = (char *)(((long)pstr + 3) & ~3);
-    dump_bytes( outfile, (char *)pModule, (int)pstr - (int)pModule, "Module", 0 );
-    return (int)pstr - (int)pModule;
+/*******************************************************************
+ *         output_bytes
+ */
+static void output_bytes( FILE *outfile, const void *buffer, unsigned int size )
+{
+    unsigned int i;
+    const unsigned char *ptr = buffer;
+
+    fprintf( outfile, "  {" );
+    for (i = 0; i < size; i++)
+    {
+        if (!(i & 7)) fprintf( outfile, "\n   " );
+        fprintf( outfile, " 0x%02x", *ptr++ );
+        if (i < size - 1) fprintf( outfile, "," );
+    }
+    fprintf( outfile, "\n  },\n" );
 }
 
 
@@ -510,25 +373,12 @@ static void output_stub_funcs( FILE *outfile, const DLLSPEC *spec )
         fprintf( outfile, "static void __wine_unimplemented( const char *func ) __attribute__((noreturn));\n" );
         fprintf( outfile, "#endif\n" );
         fprintf( outfile, "static void __wine_unimplemented( const char *func )\n{\n" );
-        fprintf( outfile, "  struct exc_record {\n" );
-        fprintf( outfile, "    unsigned int code, flags;\n" );
-        fprintf( outfile, "    void *rec, *addr;\n" );
-        fprintf( outfile, "    unsigned int params;\n" );
-        fprintf( outfile, "    const void *info[15];\n" );
-        fprintf( outfile, "  } rec;\n\n" );
-        fprintf( outfile, "  extern void __stdcall RtlRaiseException( struct exc_record * );\n\n" );
-        fprintf( outfile, "  rec.code    = 0x%08x;\n", EXCEPTION_WINE_STUB );
-        fprintf( outfile, "  rec.flags   = %d;\n", EH_NONCONTINUABLE );
-        fprintf( outfile, "  rec.rec     = 0;\n" );
-        fprintf( outfile, "  rec.params  = 2;\n" );
-        fprintf( outfile, "  rec.info[0] = \"%s\";\n", spec->file_name );
-        fprintf( outfile, "  rec.info[1] = func;\n" );
-        fprintf( outfile, "#ifdef __GNUC__\n" );
-        fprintf( outfile, "  rec.addr = __builtin_return_address(1);\n" );
-        fprintf( outfile, "#else\n" );
-        fprintf( outfile, "  rec.addr = 0;\n" );
-        fprintf( outfile, "#endif\n" );
-        fprintf( outfile, "  for (;;) RtlRaiseException( &rec );\n}\n\n" );
+        fprintf( outfile, "  extern void __stdcall RaiseException( unsigned int, unsigned int, unsigned int, const void ** );\n" );
+        fprintf( outfile, "  const void *args[2];\n" );
+        fprintf( outfile, "  args[0] = \"%s\";\n", spec->file_name );
+        fprintf( outfile, "  args[1] = func;\n" );
+        fprintf( outfile, "  for (;;) RaiseException( 0x%08x, %d, 2, args );\n}\n\n",
+                 EXCEPTION_WINE_STUB, EH_NONCONTINUABLE );
         break;
     }
     for (i = 0; i <= spec->limit; i++)
@@ -554,8 +404,17 @@ void BuildSpec16File( FILE *outfile, DLLSPEC *spec )
 {
     ORDDEF **type, **typelist;
     int i, nFuncs, nTypes;
-    int code_offset, data_offset, module_size, res_size;
-    unsigned char *data;
+    unsigned char *resdir_buffer, *resdata_buffer, *et_buffer, *data_buffer;
+    unsigned char string[256];
+    unsigned int ne_offset, segtable_offset, impnames_offset;
+    unsigned int entrypoint_size, callfrom_size;
+    unsigned int code_size, code_offset;
+    unsigned int data_size, data_offset;
+    unsigned int resnames_size, resnames_offset;
+    unsigned int resdir_size, resdir_offset;
+    unsigned int resdata_size, resdata_offset, resdata_align;
+    unsigned int et_size, et_offset;
+
     char constructor[100], destructor[100];
 #ifdef __i386__
     unsigned short code_selector = get_cs();
@@ -569,9 +428,9 @@ void BuildSpec16File( FILE *outfile, DLLSPEC *spec )
     fprintf( outfile, "extern void __wine_call_from_16_regs();\n" );
     fprintf( outfile, "extern void __wine_call_from_16_thunk();\n" );
 
-    data = (unsigned char *)xmalloc( 0x10000 );
-    memset( data, 0, 16 );
-    data_offset = 16;
+    data_buffer = xmalloc( 0x10000 );
+    memset( data_buffer, 0, 16 );
+    data_size = 16;
 
     if (!spec->dll_name)  /* set default name from file name */
     {
@@ -624,28 +483,160 @@ void BuildSpec16File( FILE *outfile, DLLSPEC *spec )
     }
 #endif
 
-    /* Output the DLL functions prototypes */
+    /* compute code and data sizes, set offsets, and output prototypes */
+
+#ifdef __i386__
+    entrypoint_size = 2 + 5 + 4;    /* pushw bp + pushl target + call */
+    callfrom_size = 5 + 7 + 4 + 8;  /* pushl relay + lcall cs:glue + lret n + args */
+#else
+    entrypoint_size = 4 + 4;  /* target + call */
+    callfrom_size = 4 + 8;    /* lret n + args */
+#endif
+    code_size = nTypes * callfrom_size;
 
     for (i = 0; i <= spec->limit; i++)
     {
         ORDDEF *odp = spec->ordinals[i];
         if (!odp) continue;
-        switch(odp->type)
+        switch (odp->type)
         {
+        case TYPE_ABS:
+            odp->offset = LOWORD(odp->u.abs.value);
+            break;
+        case TYPE_VARIABLE:
+            odp->offset = data_size;
+            memcpy( data_buffer + data_size, odp->u.var.values, odp->u.var.n_values * sizeof(int) );
+            data_size += odp->u.var.n_values * sizeof(int);
+            break;
         case TYPE_CDECL:
         case TYPE_PASCAL:
         case TYPE_VARARGS:
             fprintf( outfile, "extern void %s();\n", odp->link_name );
+            /* fall through */
+        case TYPE_STUB:
+            odp->offset = code_size;
+            code_size += entrypoint_size;
             break;
         default:
+            assert(0);
             break;
         }
     }
+    data_buffer = xrealloc( data_buffer, data_size );  /* free unneeded data */
 
-    /* Output code segment */
+    /* Output the module structure */
+
+    /* DOS header */
 
     fprintf( outfile, "\n#include \"pshpack1.h\"\n" );
-    fprintf( outfile, "\nstatic struct code_segment\n{\n" );
+    fprintf( outfile, "static const struct module_data\n{\n" );
+    fprintf( outfile, "  struct\n  {\n" );
+    fprintf( outfile, "    unsigned short e_magic;\n" );
+    fprintf( outfile, "    unsigned short e_cblp;\n" );
+    fprintf( outfile, "    unsigned short e_cp;\n" );
+    fprintf( outfile, "    unsigned short e_crlc;\n" );
+    fprintf( outfile, "    unsigned short e_cparhdr;\n" );
+    fprintf( outfile, "    unsigned short e_minalloc;\n" );
+    fprintf( outfile, "    unsigned short e_maxalloc;\n" );
+    fprintf( outfile, "    unsigned short e_ss;\n" );
+    fprintf( outfile, "    unsigned short e_sp;\n" );
+    fprintf( outfile, "    unsigned short e_csum;\n" );
+    fprintf( outfile, "    unsigned short e_ip;\n" );
+    fprintf( outfile, "    unsigned short e_cs;\n" );
+    fprintf( outfile, "    unsigned short e_lfarlc;\n" );
+    fprintf( outfile, "    unsigned short e_ovno;\n" );
+    fprintf( outfile, "    unsigned short e_res[4];\n" );
+    fprintf( outfile, "    unsigned short e_oemid;\n" );
+    fprintf( outfile, "    unsigned short e_oeminfo;\n" );
+    fprintf( outfile, "    unsigned short e_res2[10];\n" );
+    fprintf( outfile, "    unsigned int   e_lfanew;\n" );
+    fprintf( outfile, "  } dos_header;\n" );
+
+    /* NE header */
+
+    ne_offset = 64;
+    fprintf( outfile, "  struct\n  {\n" );
+    fprintf( outfile, "    unsigned short  ne_magic;\n" );
+    fprintf( outfile, "    unsigned char   ne_ver;\n" );
+    fprintf( outfile, "    unsigned char   ne_rev;\n" );
+    fprintf( outfile, "    unsigned short  ne_enttab;\n" );
+    fprintf( outfile, "    unsigned short  ne_cbenttab;\n" );
+    fprintf( outfile, "    int             ne_crc;\n" );
+    fprintf( outfile, "    unsigned short  ne_flags;\n" );
+    fprintf( outfile, "    unsigned short  ne_autodata;\n" );
+    fprintf( outfile, "    unsigned short  ne_heap;\n" );
+    fprintf( outfile, "    unsigned short  ne_stack;\n" );
+    fprintf( outfile, "    unsigned int    ne_csip;\n" );
+    fprintf( outfile, "    unsigned int    ne_sssp;\n" );
+    fprintf( outfile, "    unsigned short  ne_cseg;\n" );
+    fprintf( outfile, "    unsigned short  ne_cmod;\n" );
+    fprintf( outfile, "    unsigned short  ne_cbnrestab;\n" );
+    fprintf( outfile, "    unsigned short  ne_segtab;\n" );
+    fprintf( outfile, "    unsigned short  ne_rsrctab;\n" );
+    fprintf( outfile, "    unsigned short  ne_restab;\n" );
+    fprintf( outfile, "    unsigned short  ne_modtab;\n" );
+    fprintf( outfile, "    unsigned short  ne_imptab;\n" );
+    fprintf( outfile, "    unsigned int    ne_nrestab;\n" );
+    fprintf( outfile, "    unsigned short  ne_cmovent;\n" );
+    fprintf( outfile, "    unsigned short  ne_align;\n" );
+    fprintf( outfile, "    unsigned short  ne_cres;\n" );
+    fprintf( outfile, "    unsigned char   ne_exetyp;\n" );
+    fprintf( outfile, "    unsigned char   ne_flagsothers;\n" );
+    fprintf( outfile, "    unsigned short  ne_pretthunks;\n" );
+    fprintf( outfile, "    unsigned short  ne_psegrefbytes;\n" );
+    fprintf( outfile, "    unsigned short  ne_swaparea;\n" );
+    fprintf( outfile, "    unsigned short  ne_expver;\n" );
+    fprintf( outfile, "  } os2_header;\n" );
+
+    /* segment table */
+
+    segtable_offset = 64;
+    fprintf( outfile, "  struct\n  {\n" );
+    fprintf( outfile, "    unsigned short filepos;\n" );
+    fprintf( outfile, "    unsigned short size;\n" );
+    fprintf( outfile, "    unsigned short flags;\n" );
+    fprintf( outfile, "    unsigned short minsize;\n" );
+    fprintf( outfile, "  } segtable[2];\n" );
+
+    /* resource directory */
+
+    resdir_offset = segtable_offset + 2 * 8;
+    resdir_size = get_res16_directory_size( spec );
+    fprintf( outfile, "  unsigned char resdir[%d];\n", resdir_size );
+
+    /* resident names table */
+
+    resnames_offset = resdir_offset + resdir_size;
+    fprintf( outfile, "  struct\n  {\n" );
+    fprintf( outfile, "    struct { unsigned char len; char name[%d]; unsigned short ord; } name_0;\n",
+             strlen( spec->dll_name ) );
+    resnames_size = 3 + strlen( spec->dll_name );
+    for (i = 1; i <= spec->limit; i++)
+    {
+        ORDDEF *odp = spec->ordinals[i];
+        if (!odp || !odp->name[0]) continue;
+        fprintf( outfile, "    struct { unsigned char len; char name[%d]; unsigned short ord; } name_%d;\n",
+                 strlen(odp->name), i );
+        resnames_size += 3 + strlen( odp->name );
+    }
+    fprintf( outfile, "    unsigned char name_last[%d];\n", 2 - (resnames_size & 1) );
+    resnames_size = (resnames_size + 2) & ~1;
+    fprintf( outfile, "  } resnames;\n" );
+
+    /* imported names table */
+
+    impnames_offset = resnames_offset + resnames_size;
+    fprintf( outfile, "  unsigned char impnames[2];\n" );
+
+    /* entry table */
+
+    et_offset = impnames_offset + 2;
+    et_size = output_entry_table( &et_buffer, spec );
+    fprintf( outfile, "  unsigned char entry_table[%d];\n", et_size );
+
+    /* code segment */
+
+    code_offset = et_offset + et_size;
     fprintf( outfile, "  struct {\n" );
 #ifdef __i386__
     fprintf( outfile, "    unsigned char pushl;\n" );      /* pushl $relay */
@@ -667,10 +658,110 @@ void BuildSpec16File( FILE *outfile, DLLSPEC *spec )
     fprintf( outfile, "    unsigned short call;\n" );      /* call CALLFROM16 */
     fprintf( outfile, "    short callfrom16;\n" );
     fprintf( outfile, "  } entry[%d];\n", nFuncs );
-    fprintf( outfile, "} code_segment =\n{\n  {\n" );
 
-    code_offset = 0;
+    /* data segment */
 
+    data_offset = code_offset + code_size;
+    fprintf( outfile, "  unsigned char data_segment[%d];\n", data_size );
+    if (data_offset + data_size >= 0x10000)
+        fatal_error( "Not supported yet: 16-bit module data larger than 64K\n" );
+
+    /* resource data */
+
+    resdata_offset = ne_offset + data_offset + data_size;
+    for (resdata_align = 0; resdata_align < 16; resdata_align++)
+    {
+        unsigned int size = get_res16_data_size( spec, resdata_offset, resdata_align );
+        if ((resdata_offset + size) >> resdata_align <= 0xffff) break;
+    }
+    output_res16_directory( &resdir_buffer, spec, resdata_offset, resdata_align );
+    resdata_size = output_res16_data( &resdata_buffer, spec, resdata_offset, resdata_align );
+    if (resdata_size) fprintf( outfile, "  unsigned char resources[%d];\n", resdata_size );
+
+    /* Output the module data */
+
+    /* DOS header */
+
+    fprintf( outfile, "} module =\n{\n  {\n" );
+    fprintf( outfile, "    0x%04x,\n", IMAGE_DOS_SIGNATURE );  /* e_magic */
+    fprintf( outfile, "    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n" );
+    fprintf( outfile, "    { 0, 0, 0, 0, }, 0, 0,\n" );
+    fprintf( outfile, "    { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },\n" );
+    fprintf( outfile, "    sizeof(module.dos_header)\n" );  /* e_lfanew */
+
+    /* NE header */
+
+    fprintf( outfile, "  },\n  {\n" );
+    fprintf( outfile, "    0x%04x,\n", IMAGE_OS2_SIGNATURE ); /* ne_magic */
+    fprintf( outfile, "    0, 0,\n" );
+    fprintf( outfile, "    %d,\n", et_offset );               /* ne_enttab */
+    fprintf( outfile, "    sizeof(module.entry_table),\n" );  /* ne_cbenttab */
+    fprintf( outfile, "    0,\n" );                           /* ne_crc */
+    fprintf( outfile, "    0x%04x,\n",                        /* ne_flags */
+             NE_FFLAGS_SINGLEDATA | NE_FFLAGS_LIBMODULE );
+    fprintf( outfile, "    2,\n" );                           /* ne_autodata */
+    fprintf( outfile, "    %d,\n", spec->heap_size );         /* ne_heap */
+    fprintf( outfile, "    0, 0, 0,\n" );
+    fprintf( outfile, "    2,\n" );                           /* ne_cseg */
+    fprintf( outfile, "    0,\n" );                           /* ne_cmod */
+    fprintf( outfile, "    0,\n" );                           /* ne_cbnrestab */
+    fprintf( outfile, "    %d,\n", segtable_offset );         /* ne_segtab */
+    fprintf( outfile, "    %d,\n", resdir_offset );           /* ne_rsrctab */
+    fprintf( outfile, "    %d,\n", resnames_offset );         /* ne_restab */
+    fprintf( outfile, "    %d,\n", impnames_offset );         /* ne_modtab */
+    fprintf( outfile, "    %d,\n", impnames_offset );         /* ne_imptab */
+    fprintf( outfile, "    0,\n" );                           /* ne_nrestab */
+    fprintf( outfile, "    0,\n" );                           /* ne_cmovent */
+    fprintf( outfile, "    0,\n" );                           /* ne_align */
+    fprintf( outfile, "    0,\n" );                           /* ne_cres */
+    fprintf( outfile, "    0x%04x,\n", NE_OSFLAGS_WINDOWS );  /* ne_exetyp */
+    fprintf( outfile, "    0x%04x,\n", NE_AFLAGS_FASTLOAD );  /* ne_flagsothers */
+    fprintf( outfile, "    0,\n" );                           /* ne_pretthunks */
+    fprintf( outfile, "    0,\n" );                           /* ne_psegrefbytes */
+    fprintf( outfile, "    0,\n" );                           /* ne_swaparea */
+    fprintf( outfile, "    0\n" );                            /* ne_expver */
+    fprintf( outfile, "  },\n" );
+
+    /* segment table */
+
+    fprintf( outfile, "  {\n" );
+    fprintf( outfile, "    { %d, %d, 0x%04x, %d },\n",
+             ne_offset + code_offset, code_size, NE_SEGFLAGS_32BIT, code_size );
+    fprintf( outfile, "    { %d, %d, 0x%04x, %d },\n",
+             ne_offset + data_offset, data_size, NE_SEGFLAGS_DATA, data_size );
+    fprintf( outfile, "  },\n" );
+
+    /* resource directory */
+
+    output_bytes( outfile, resdir_buffer, resdir_size );
+    free( resdir_buffer );
+
+    /* resident names table */
+
+    fprintf( outfile, "  {\n" );
+    strcpy( string, spec->dll_name );
+    fprintf( outfile, "    { %d, \"%s\", 0 },\n", strlen(string), strupper(string) );
+    for (i = 1; i <= spec->limit; i++)
+    {
+        ORDDEF *odp = spec->ordinals[i];
+        if (!odp || !odp->name[0]) continue;
+        strcpy( string, odp->name );
+        fprintf( outfile, "    { %d, \"%s\", %d },\n", strlen(string), strupper(string), i );
+    }
+    fprintf( outfile, "    { 0 }\n  },\n" );
+
+    /* imported names table */
+
+    fprintf( outfile, "  { 0, 0 },\n" );
+
+    /* entry table */
+
+    output_bytes( outfile, et_buffer, et_size );
+    free( et_buffer );
+
+    /* code segment */
+
+    fprintf( outfile, "  {\n" );
     for ( i = 0; i < nTypes; i++ )
     {
         char profile[101], *arg;
@@ -732,7 +823,6 @@ void BuildSpec16File( FILE *outfile, DLLSPEC *spec )
             fprintf( outfile, "     { 0xcb66, 0x9090, { 0x%08x, 0x%08x } },\n",
                      arg_types[0], arg_types[1] );
 #endif
-        code_offset += sizeof(CALLFROM16);
     }
     fprintf( outfile, "  },\n  {\n" );
 
@@ -742,15 +832,6 @@ void BuildSpec16File( FILE *outfile, DLLSPEC *spec )
         if (!odp) continue;
         switch (odp->type)
         {
-          case TYPE_ABS:
-            odp->offset = LOWORD(odp->u.abs.value);
-            break;
-
-          case TYPE_VARIABLE:
-            odp->offset = data_offset;
-            data_offset += StoreVariableCode( data + data_offset, 4, odp);
-            break;
-
           case TYPE_CDECL:
           case TYPE_PASCAL:
           case TYPE_VARARGS:
@@ -759,57 +840,38 @@ void BuildSpec16File( FILE *outfile, DLLSPEC *spec )
             assert( type );
 
             fprintf( outfile, "    /* %s.%d */ ", spec->dll_name, i );
+            fprintf( outfile,
 #ifdef __i386__
-            fprintf( outfile, "{ 0x5566, 0x68, %s, 0xe866, %d  /* %s */ },\n",
+                     "{ 0x5566, 0x68, %s, 0xe866, %d  /* %s */ },\n",
 #else
-            fprintf( outfile, "{ %s, 0xe866, %d, /* %s */ },\n",
+                     "{ %s, 0xe866, %d, /* %s */ },\n",
 #endif
                      odp->link_name,
-                     (type-typelist)*sizeof(CALLFROM16) -
-                     (code_offset + sizeof(ENTRYPOINT16)),
+                     (type - typelist) * callfrom_size - (odp->offset + entrypoint_size),
                      get_function_name( odp ) );
-
-            odp->offset = code_offset;
-            code_offset += sizeof(ENTRYPOINT16);
             break;
-
-          default:
-            fprintf(stderr,"build: function type %d not available for Win16\n",
-                    odp->type);
-            exit(1);
+        default:
+            break;
         }
     }
+    fprintf( outfile, "  },\n" );
 
-    fprintf( outfile, "    }\n};\n" );
 
-    /* Output data segment */
+    /* data_segment */
 
-    dump_bytes( outfile, data, data_offset, "Data_Segment", 0 );
+    output_bytes( outfile, data_buffer, data_size );
+    free( data_buffer );
 
-    /* Build the module */
+    /* resource data */
 
-    module_size = BuildModule16( outfile, code_offset, data_offset, spec );
-    res_size = output_res16_data( outfile, spec );
+    if (resdata_size)
+    {
+        output_bytes( outfile, resdata_buffer, resdata_size );
+        free( resdata_buffer );
+    }
 
-    /* Output the DLL descriptor */
-
-    fprintf( outfile, "#include \"poppack.h\"\n\n" );
-
-    fprintf( outfile, "static const struct dll_descriptor\n{\n" );
-    fprintf( outfile, "    unsigned char *module_start;\n" );
-    fprintf( outfile, "    int module_size;\n" );
-    fprintf( outfile, "    struct code_segment *code_start;\n" );
-    fprintf( outfile, "    unsigned char *data_start;\n" );
-    fprintf( outfile, "    const char *owner;\n" );
-    fprintf( outfile, "    const unsigned char *rsrc;\n" );
-    fprintf( outfile, "} descriptor =\n{\n" );
-    fprintf( outfile, "    Module,\n" );
-    fprintf( outfile, "    sizeof(Module),\n" );
-    fprintf( outfile, "    &code_segment,\n" );
-    fprintf( outfile, "    Data_Segment,\n" );
-    fprintf( outfile, "    \"%s\",\n", spec->owner_name );
-    fprintf( outfile, "    %s\n", res_size ? "resource_data" : "0" );
     fprintf( outfile, "};\n" );
+    fprintf( outfile, "#include \"poppack.h\"\n\n" );
 
     /* Output the DLL constructor */
 
@@ -820,13 +882,13 @@ void BuildSpec16File( FILE *outfile, DLLSPEC *spec )
     fprintf( outfile,
              "void %s(void)\n"
              "{\n"
-             "    extern void __wine_register_dll_16( const struct dll_descriptor *descr );\n"
-             "    __wine_register_dll_16( &descriptor );\n"
-             "}\n", constructor );
+             "    extern void __wine_dll_register_16( const struct module_data *, const char * );\n"
+             "    __wine_dll_register_16( &module, \"%s\" );\n"
+             "}\n", constructor, spec->file_name );
     fprintf( outfile,
              "void %s(void)\n"
              "{\n"
-             "    extern void __wine_unregister_dll_16( const struct dll_descriptor *descr );\n"
-             "    __wine_unregister_dll_16( &descriptor );\n"
+             "    extern void __wine_dll_unregister_16( const struct module_data * );\n"
+             "    __wine_dll_unregister_16( &module );\n"
              "}\n", destructor );
 }
diff --git a/reactos/tools/winebuild/spec32.c b/reactos/tools/winebuild/spec32.c
index 6fb1f4ae0ef..6c3e1ee88dd 100644
--- a/reactos/tools/winebuild/spec32.c
+++ b/reactos/tools/winebuild/spec32.c
@@ -176,9 +176,7 @@ static int output_exports( FILE *outfile, int nr_exports, DLLSPEC *spec )
         case TYPE_CDECL:
             if (!(odp->flags & FLAG_FORWARD))
             {
-                fprintf( outfile, "    \"\\t.long " __ASM_NAME("%s") "\\n\"\n",
-                         (odp->flags & FLAG_REGISTER) ? make_internal_name( odp, spec, "regs" )
-                         : odp->link_name );
+                fprintf( outfile, "    \"\\t.long " __ASM_NAME("%s") "\\n\"\n", odp->link_name );
             }
             else
             {
@@ -261,7 +259,6 @@ static int output_exports( FILE *outfile, int nr_exports, DLLSPEC *spec )
         {
             ORDDEF *odp = spec->ordinals[i];
             unsigned int j, args, mask = 0;
-            const char *name;
 
             /* skip nonexistent entry points */
             if (!odp) goto ignore;
@@ -277,22 +274,20 @@ static int output_exports( FILE *outfile, int nr_exports, DLLSPEC *spec )
             }
             if ((odp->flags & FLAG_RET64) && (j < 16)) mask |= 0x80000000;
 
-            name = odp->link_name;
             args = strlen(odp->u.func.arg_types) * sizeof(int);
-            if (odp->flags & FLAG_REGISTER) name = make_internal_name( odp, spec, "regs" );
 
             switch(odp->type)
             {
             case TYPE_STDCALL:
-                fprintf( outfile, "    \"\\tjmp " __ASM_NAME("%s") "\\n\"\n", name );
+                fprintf( outfile, "    \"\\tjmp " __ASM_NAME("%s") "\\n\"\n", odp->link_name );
                 fprintf( outfile, "    \"\\tret $%d\\n\"\n", args );
-                fprintf( outfile, "    \"\\t.long " __ASM_NAME("%s") ",0x%08x\\n\"\n", name, mask );
+                fprintf( outfile, "    \"\\t.long " __ASM_NAME("%s") ",0x%08x\\n\"\n", odp->link_name, mask );
                 break;
             case TYPE_CDECL:
-                fprintf( outfile, "    \"\\tjmp " __ASM_NAME("%s") "\\n\"\n", name );
+                fprintf( outfile, "    \"\\tjmp " __ASM_NAME("%s") "\\n\"\n", odp->link_name );
                 fprintf( outfile, "    \"\\tret\\n\"\n" );
                 fprintf( outfile, "    \"\\t" __ASM_SHORT " %d\\n\"\n", args );
-                fprintf( outfile, "    \"\\t.long " __ASM_NAME("%s") ",0x%08x\\n\"\n", name, mask );
+                fprintf( outfile, "    \"\\t.long " __ASM_NAME("%s") ",0x%08x\\n\"\n", odp->link_name, mask );
                 break;
             default:
                 assert(0);
@@ -367,38 +362,6 @@ static void output_stub_funcs( FILE *outfile, DLLSPEC *spec )
 }
 
 
-/*******************************************************************
- *         output_register_funcs
- *
- * Output the functions for register entry points
- */
-static void output_register_funcs( FILE *outfile, DLLSPEC *spec )
-{
-    const char *name;
-    int i;
-
-    for (i = 0; i < spec->nb_entry_points; i++)
-    {
-        const ORDDEF *odp = &spec->entry_points[i];
-        if (odp->type != TYPE_STDCALL && odp->type != TYPE_CDECL) continue;
-        if (!(odp->flags & FLAG_REGISTER)) continue;
-        if (odp->flags & FLAG_FORWARD) continue;
-        name = make_internal_name( odp, spec, "regs" );
-        fprintf( outfile,
-                 "asm(\".align %d\\n\\t\"\n"
-                 "    \"" __ASM_FUNC("%s") "\\n\\t\"\n"
-                 "    \"" __ASM_NAME("%s") ":\\n\\t\"\n"
-                 "    \"call " __ASM_NAME("__wine_call_from_32_regs") "\\n\\t\"\n"
-                 "    \".long " __ASM_NAME("%s") "\\n\\t\"\n"
-                 "    \".byte %d,%d\");\n",
-                 get_alignment(4),
-                 name, name, odp->link_name,
-                 strlen(odp->u.func.arg_types) * sizeof(int),
-                 (odp->type == TYPE_CDECL) ? 0 : (strlen(odp->u.func.arg_types) * sizeof(int)) );
-    }
-}
-
-
 /*******************************************************************
  *         output_dll_init
  *
@@ -499,7 +462,7 @@ void output_dll_init( FILE *outfile, const char *constructor, const char *destru
 void BuildSpec32File( FILE *outfile, DLLSPEC *spec )
 {
     int exports_size = 0;
-    int nr_exports, nr_imports;
+    int nr_exports, nr_imports, nr_delayed;
     DWORD page_size;
     const char *init_func = spec->init_func;
 
@@ -561,10 +524,6 @@ void BuildSpec32File( FILE *outfile, DLLSPEC *spec )
         fprintf( outfile, "static void __asm__dummy(void) {\n" );
         fprintf( outfile, "#endif /* !defined(__GNUC__) */\n" );
 
-        /* Output code for all register functions */
-
-        output_register_funcs( outfile, spec );
-
         /* Output the exports and relay entry points */
 
         exports_size = output_exports( outfile, nr_exports, spec );
@@ -576,7 +535,7 @@ void BuildSpec32File( FILE *outfile, DLLSPEC *spec )
 
     /* Output the DLL imports */
 
-    nr_imports = output_imports( outfile, spec );
+    nr_imports = output_imports( outfile, spec, &nr_delayed );
 
     /* Output the resources */
 
@@ -626,10 +585,20 @@ void BuildSpec32File( FILE *outfile, DLLSPEC *spec )
                  "    if (reason == %d && __wine_spec_init_state == 1)\n"
                  "        _init( __wine_main_argc, __wine_main_argv, __wine_main_environ );\n"
                  "    ret = %s ? %s( inst, reason, reserved ) : 1;\n"
-                 "    if (reason == %d && __wine_spec_init_state == 1) _fini();\n"
-                 "    return ret;\n"
-                 "}\n",
+                 "    if (reason == %d && __wine_spec_init_state == 1)\n",
                  DLL_PROCESS_ATTACH, init_func, init_func, DLL_PROCESS_DETACH );
+        if (!nr_delayed)
+            fprintf( outfile, "        _fini();\n" );
+        else
+            fprintf( outfile,
+                     "    {\n"
+                     "        extern int __stdcall FreeLibrary(void *);\n"
+                     "        unsigned int i;\n"
+                     "        _fini();\n"
+                     "        for (i = 0; i < sizeof(__wine_delay_imp_hmod)/sizeof(__wine_delay_imp_hmod[0]); i++)\n"
+                     "            if (__wine_delay_imp_hmod[i]) FreeLibrary( __wine_delay_imp_hmod[i] );\n"
+                     "    }\n" );
+        fprintf( outfile, "    return ret;\n}\n" );
         init_func = "__wine_dll_main";
     }
     else switch(spec->subsystem)
diff --git a/reactos/tools/winebuild/utils.c b/reactos/tools/winebuild/utils.c
index 79b3899d97e..1fe09aa30f8 100644
--- a/reactos/tools/winebuild/utils.c
+++ b/reactos/tools/winebuild/utils.c
@@ -74,6 +74,13 @@ char *strupper(char *s)
     return s;
 }
 
+int strendswith(const char* str, const char* end)
+{
+    int l = strlen(str);
+    int m = strlen(end);
+    return l >= m && strcmp(str + l - m, end) == 0;
+}
+
 void fatal_error( const char *msg, ... )
 {
     va_list valist;
diff --git a/reactos/tools/winebuild/winebuild.man.in b/reactos/tools/winebuild/winebuild.man.in
index cca9dcf34c9..19fddf5a648 100644
--- a/reactos/tools/winebuild/winebuild.man.in
+++ b/reactos/tools/winebuild/winebuild.man.in
@@ -69,6 +69,12 @@ Change to the specified directory before reading source files. Only
 meaningful in
 .BR \--debug\  mode.
 .TP
+.BI \-d,\ --delay-lib= name
+Set the delayed import mode for the specified library, which must be
+one of the libraries imported with the \fB-l\fR option. Delayed mode
+means that the library won't be loaded until a function imported from
+it is actually called.
+.TP
 .BI \-D\  symbol
 Ignored for compatibility with the C compiler.
 .TP
@@ -138,11 +144,6 @@ Import the specified library, looking for a corresponding
 \fIlibname.def\fR file in the directories specified with the \fB-L\fR
 option.
 .TP
-.BI \-d,\ --delay-lib= name
-Same as the \fB-l\fR option, but import the specified library in
-delayed mode (i.e. the library won't be loaded until a function
-imported from it is actually called).
-.TP
 .BI \-M,\ --main-module= module
 Specify that we are building a 16-bit dll, that will ultimately be
 linked together with the 32-bit dll specified in \fImodule\fR.  Only
@@ -249,7 +250,7 @@ The function returns a 64-bit value (Win32 only).
 The entry point is only available on i386 platforms.
 .TP
 .B -register
-The function uses CPU register to pass arguments.
+The function uses CPU register to pass arguments (Win16 only).
 .TP
 .B -private
 The function cannot be imported from other dlls, it can only be