diff --git a/reactos/tools/winebuild/build.h b/reactos/tools/winebuild/build.h
index 8543ef6067e..7f66b5b8b42 100644
--- a/reactos/tools/winebuild/build.h
+++ b/reactos/tools/winebuild/build.h
@@ -189,6 +189,7 @@ extern const char *asm_globl( const char *func );
 extern const char *get_asm_ptr_keyword(void);
 extern const char *get_asm_string_keyword(void);
 extern const char *get_asm_short_keyword(void);
+extern const char *get_asm_rodata_section(void);
 extern const char *get_asm_string_section(void);
 extern void output_function_size( FILE *outfile, const char *name );
 
@@ -199,6 +200,7 @@ extern void add_extra_ld_symbol( const char *name );
 extern void read_undef_symbols( DLLSPEC *spec, char **argv );
 extern int resolve_imports( DLLSPEC *spec );
 extern int has_imports(void);
+extern int has_relays( DLLSPEC *spec );
 extern void output_get_pc_thunk( FILE *outfile );
 extern void output_stubs( FILE *outfile, DLLSPEC *spec );
 extern void output_imports( FILE *outfile, DLLSPEC *spec );
diff --git a/reactos/tools/winebuild/import.c b/reactos/tools/winebuild/import.c
index 6f8bd28840f..a51eb30cbce 100644
--- a/reactos/tools/winebuild/import.c
+++ b/reactos/tools/winebuild/import.c
@@ -1245,5 +1245,6 @@ void output_imports( FILE *outfile, DLLSPEC *spec )
     output_immediate_import_thunks( outfile );
     output_delayed_import_thunks( outfile, spec );
     output_external_link_imports( outfile, spec );
-    if (nb_imports || ext_link_imports.count || has_stubs(spec)) output_get_pc_thunk( outfile );
+    if (nb_imports || ext_link_imports.count || has_stubs(spec) || has_relays(spec))
+        output_get_pc_thunk( outfile );
 }
diff --git a/reactos/tools/winebuild/parser.c b/reactos/tools/winebuild/parser.c
index 5352b7a2ec7..1ea33766523 100644
--- a/reactos/tools/winebuild/parser.c
+++ b/reactos/tools/winebuild/parser.c
@@ -516,12 +516,6 @@ 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 e53a8392b7d..ccf37ec78d3 100644
--- a/reactos/tools/winebuild/relay.c
+++ b/reactos/tools/winebuild/relay.c
@@ -849,7 +849,7 @@ static void BuildCallFrom32Regs( FILE *outfile )
     output_function_size( outfile, "__wine_call_from_32_regs" );
 
     function_header( outfile, "__wine_call_from_32_restore_regs" );
-    fprintf( outfile, "\tleal 4(%%esp),%%ecx\n" );
+    fprintf( outfile, "\tmovl 4(%%esp),%%ecx\n" );
     fprintf( outfile, "\tjmp 2b\n" );
     output_function_size( outfile, "__wine_call_from_32_restore_regs" );
 }
diff --git a/reactos/tools/winebuild/spec32.c b/reactos/tools/winebuild/spec32.c
index ed500c7b176..bf45ffdb469 100644
--- a/reactos/tools/winebuild/spec32.c
+++ b/reactos/tools/winebuild/spec32.c
@@ -37,7 +37,32 @@
 #include "build.h"
 
 
+/* check if entry point needs a relay thunk */
+static inline int needs_relay( const ORDDEF *odp )
+{
+    /* skip nonexistent entry points */
+    if (!odp) return 0;
+    /* skip non-functions */
+    if ((odp->type != TYPE_STDCALL) && (odp->type != TYPE_CDECL)) return 0;
+    /* skip norelay and forward entry points */
+    if (odp->flags & (FLAG_NORELAY|FLAG_FORWARD)) return 0;
+    return 1;
+}
 
+/* check if dll will output relay thunks */
+int has_relays( DLLSPEC *spec )
+{
+    unsigned int i;
+
+    if (target_cpu != CPU_x86) return 0;
+
+    for (i = spec->base; i <= spec->limit; i++)
+    {
+        ORDDEF *odp = spec->ordinals[i];
+        if (needs_relay( odp )) return 1;
+    }
+    return 0;
+}
 
 /*******************************************************************
  *         make_internal_name
@@ -61,6 +86,101 @@ static const char *make_internal_name( const ORDDEF *odp, DLLSPEC *spec, const c
 }
 
 
+/*******************************************************************
+ *         output_relay_debug
+ *
+ * Output entry points for relay debugging
+ */
+static void output_relay_debug( FILE *outfile, DLLSPEC *spec )
+{
+    unsigned int i, j, args, flags;
+
+    /* first the table of entry point offsets */
+
+    fprintf( outfile, "\t%s\n", get_asm_rodata_section() );
+    fprintf( outfile, "\t.align %d\n", get_alignment(4) );
+    fprintf( outfile, ".L__wine_spec_relay_entry_point_offsets:\n" );
+
+    for (i = spec->base; i <= spec->limit; i++)
+    {
+        ORDDEF *odp = spec->ordinals[i];
+
+        if (needs_relay( odp ))
+            fprintf( outfile, "\t.long .L__wine_spec_relay_entry_point_%d-__wine_spec_relay_entry_points\n", i );
+        else
+            fprintf( outfile, "\t.long 0\n" );
+    }
+
+    /* then the table of argument types */
+
+    fprintf( outfile, "\t.align %d\n", get_alignment(4) );
+    fprintf( outfile, ".L__wine_spec_relay_arg_types:\n" );
+
+    for (i = spec->base; i <= spec->limit; i++)
+    {
+        ORDDEF *odp = spec->ordinals[i];
+        unsigned int mask = 0;
+
+        if (needs_relay( odp ))
+        {
+            for (j = 0; j < 16 && odp->u.func.arg_types[j]; j++)
+            {
+                if (odp->u.func.arg_types[j] == 't') mask |= 1<< (j*2);
+                if (odp->u.func.arg_types[j] == 'W') mask |= 2<< (j*2);
+            }
+        }
+        fprintf( outfile, "\t.long 0x%08x\n", mask );
+    }
+
+    /* then the relay thunks */
+
+    fprintf( outfile, "\t.text\n" );
+    fprintf( outfile, "__wine_spec_relay_entry_points:\n" );
+    fprintf( outfile, "\tnop\n" );  /* to avoid 0 offset */
+
+    for (i = spec->base; i <= spec->limit; i++)
+    {
+        ORDDEF *odp = spec->ordinals[i];
+
+        if (!needs_relay( odp )) continue;
+
+        fprintf( outfile, "\t.align %d\n", get_alignment(4) );
+        fprintf( outfile, ".L__wine_spec_relay_entry_point_%d:\n", i );
+
+        if (odp->flags & FLAG_REGISTER)
+            fprintf( outfile, "\tpushl %%eax\n" );
+        else
+            fprintf( outfile, "\tpushl %%esp\n" );
+
+        args = strlen(odp->u.func.arg_types);
+        flags = 0;
+        if (odp->flags & FLAG_RET64) flags |= 1;
+        if (odp->type == TYPE_STDCALL) flags |= 2;
+        fprintf( outfile, "\tpushl $%u\n", (flags << 24) | (args << 16) | (i - spec->base) );
+
+        if (UsePIC)
+        {
+            fprintf( outfile, "\tcall %s\n", asm_name("__wine_spec_get_pc_thunk_eax") );
+            fprintf( outfile, "1:\tleal .L__wine_spec_relay_descr-1b(%%eax),%%eax\n" );
+        }
+        else fprintf( outfile, "\tmovl $.L__wine_spec_relay_descr,%%eax\n" );
+        fprintf( outfile, "\tpushl %%eax\n" );
+
+        if (odp->flags & FLAG_REGISTER)
+        {
+            fprintf( outfile, "\tcall *8(%%eax)\n" );
+        }
+        else
+        {
+            fprintf( outfile, "\tcall *4(%%eax)\n" );
+            if (odp->type == TYPE_STDCALL)
+                fprintf( outfile, "\tret $%u\n", args * get_ptr_size() );
+            else
+                fprintf( outfile, "\tret\n" );
+        }
+    }
+}
+
 /*******************************************************************
  *         output_exports
  *
@@ -183,58 +303,27 @@ static void output_exports( FILE *outfile, DLLSPEC *spec )
                 fprintf( outfile, "\t%s \"%s\"\n", get_asm_string_keyword(), odp->link_name );
         }
     }
-    fprintf( outfile, "\t.align %d\n", get_alignment(4) );
+    fprintf( outfile, "\t.align %d\n", get_alignment(get_ptr_size()) );
     fprintf( outfile, ".L__wine_spec_exports_end:\n" );
 
     /* output relays */
 
     /* we only support relay debugging on i386 */
-    if (target_cpu == CPU_x86)
+    if (target_cpu != CPU_x86)
     {
-        for (i = spec->base; i <= spec->limit; i++)
-        {
-            ORDDEF *odp = spec->ordinals[i];
-            unsigned int j, args, mask = 0;
-
-            /* skip nonexistent entry points */
-            if (!odp) goto ignore;
-            /* skip non-functions */
-            if ((odp->type != TYPE_STDCALL) && (odp->type != TYPE_CDECL)) goto ignore;
-            /* skip norelay and forward entry points */
-            if (odp->flags & (FLAG_NORELAY|FLAG_FORWARD)) goto ignore;
-
-            for (j = 0; odp->u.func.arg_types[j]; j++)
-            {
-                if (odp->u.func.arg_types[j] == 't') mask |= 1<< (j*2);
-                if (odp->u.func.arg_types[j] == 'W') mask |= 2<< (j*2);
-            }
-            if ((odp->flags & FLAG_RET64) && (j < 16)) mask |= 0x80000000;
-
-            args = strlen(odp->u.func.arg_types) * get_ptr_size();
-
-            switch(odp->type)
-            {
-            case TYPE_STDCALL:
-                fprintf( outfile, "\tjmp %s\n", asm_name(odp->link_name) );
-                fprintf( outfile, "\tret $%d\n", args );
-                fprintf( outfile, "\t.long %s,0x%08x\n", asm_name(odp->link_name), mask );
-                break;
-            case TYPE_CDECL:
-                fprintf( outfile, "\tjmp %s\n", asm_name(odp->link_name) );
-                fprintf( outfile, "\tret\n" );
-                fprintf( outfile, "\t%s %d\n", get_asm_short_keyword(), args );
-                fprintf( outfile, "\t.long %s,0x%08x\n", asm_name(odp->link_name), mask );
-                break;
-            default:
-                assert(0);
-            }
-            continue;
-
-        ignore:
-            fprintf( outfile, "\t.long 0,0,0,0\n" );
-        }
+        fprintf( outfile, "\t%s 0\n", get_asm_ptr_keyword() );
+        return;
     }
-    else fprintf( outfile, "\t.long 0\n" );
+
+    fprintf( outfile, ".L__wine_spec_relay_descr:\n" );
+    fprintf( outfile, "\t%s 0xdeb90001\n", get_asm_ptr_keyword() );  /* magic */
+    fprintf( outfile, "\t%s 0,0\n", get_asm_ptr_keyword() );         /* relay funcs */
+    fprintf( outfile, "\t%s 0\n", get_asm_ptr_keyword() );           /* private data */
+    fprintf( outfile, "\t%s __wine_spec_relay_entry_points\n", get_asm_ptr_keyword() );
+    fprintf( outfile, "\t%s .L__wine_spec_relay_entry_point_offsets\n", get_asm_ptr_keyword() );
+    fprintf( outfile, "\t%s .L__wine_spec_relay_arg_types\n", get_asm_ptr_keyword() );
+
+    output_relay_debug( outfile, spec );
 }
 
 
diff --git a/reactos/tools/winebuild/utils.c b/reactos/tools/winebuild/utils.c
index fbbfa455b41..85d552355a2 100644
--- a/reactos/tools/winebuild/utils.c
+++ b/reactos/tools/winebuild/utils.c
@@ -585,6 +585,15 @@ const char *get_asm_short_keyword(void)
     }
 }
 
+const char *get_asm_rodata_section(void)
+{
+    switch (target_platform)
+    {
+    case PLATFORM_APPLE: return ".const";
+    default:             return ".section .rodata";
+    }
+}
+
 const char *get_asm_string_section(void)
 {
     switch (target_platform)
diff --git a/reactos/tools/winebuild/winebuild.man.in b/reactos/tools/winebuild/winebuild.man.in
index a682a10646b..6c9bafcbdd5 100644
--- a/reactos/tools/winebuild/winebuild.man.in
+++ b/reactos/tools/winebuild/winebuild.man.in
@@ -260,7 +260,7 @@ 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 (Win16 only).
+The function uses CPU register to pass arguments.
 .TP
 .B -private
 The function cannot be imported from other dlls, it can only be