tools/winebuild/relay.c

   1 /*
   2  * Relay calls helper routines
   3  *
   4  * Copyright 1993 Robert J. Amstadt
   5  * Copyright 1995 Martin von Loewis
   6  * Copyright 1995, 1996, 1997 Alexandre Julliard
   7  * Copyright 1997 Eric Youngdale
   8  * Copyright 1999 Ulrich Weigand
   9  *
  10  * This library is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2.1 of the License, or (at your option) any later version.
  14  *
  15  * This library is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * Lesser General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU Lesser General Public
  21  * License along with this library; if not, write to the Free Software
  22  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  23  */
  24
  25 #include "config.h"
  26 #include "wine/port.h"
  27
  28 #include <ctype.h>
  29
  30 #include "thread.h"
  31 #include "stackframe.h"
  32
  33 #include "build.h"
  34
  35 #if defined(__GNUC__) && !defined(__svr4__)
  36 static const int use_stabs = 1;
  37 #else
  38 static const int use_stabs = 0;
  39 #endif
  40
  41 #ifdef __i386__
  42
  43 static void function_header( FILE *outfile, const char *name )
  44 {
  45     fprintf( outfile, "\n\t.align %d\n", get_alignment(4) );
  46     if (use_stabs)
  47         fprintf( outfile, "\t.stabs \"%s:F1\",36,0,0," __ASM_NAME("%s") "\n", name, name);
  48     fprintf( outfile, "\t" __ASM_FUNC("%s") "\n", name );
  49     fprintf( outfile, "\t.globl " __ASM_NAME("%s") "\n", name );
  50     fprintf( outfile, __ASM_NAME("%s") ":\n", name );
  51 }
  52
  53
  54 /*******************************************************************
  55  *         BuildCallFrom16Core
  56  *
  57  * This routine builds the core routines used in 16->32 thunks:
  58  * CallFrom16Word, CallFrom16Long, CallFrom16Register, and CallFrom16Thunk.
  59  *
  60  * These routines are intended to be called via a far call (with 32-bit
  61  * operand size) from 16-bit code.  The 16-bit code stub must push %bp,
  62  * the 32-bit entry point to be called, and the argument conversion
  63  * routine to be used (see stack layout below).
  64  *
  65  * The core routine completes the STACK16FRAME on the 16-bit stack and
  66  * switches to the 32-bit stack.  Then, the argument conversion routine
  67  * is called; it gets passed the 32-bit entry point and a pointer to the
  68  * 16-bit arguments (on the 16-bit stack) as parameters. (You can either
  69  * use conversion routines automatically generated by BuildCallFrom16,
  70  * or write your own for special purposes.)
  71  *
  72  * The conversion routine must call the 32-bit entry point, passing it
  73  * the converted arguments, and return its return value to the core.
  74  * After the conversion routine has returned, the core switches back
  75  * to the 16-bit stack, converts the return value to the DX:AX format
  76  * (CallFrom16Long), and returns to the 16-bit call stub.  All parameters,
  77  * including %bp, are popped off the stack.
  78  *
  79  * The 16-bit call stub now returns to the caller, popping the 16-bit
  80  * arguments if necessary (pascal calling convention).
  81  *
  82  * In the case of a 'register' function, CallFrom16Register fills a
  83  * CONTEXT86 structure with the values all registers had at the point
  84  * the first instruction of the 16-bit call stub was about to be
  85  * executed.  A pointer to this CONTEXT86 is passed as third parameter
  86  * to the argument conversion routine, which typically passes it on
  87  * to the called 32-bit entry point.
  88  *
  89  * CallFrom16Thunk is a special variant used by the implementation of
  90  * the Win95 16->32 thunk functions C16ThkSL and C16ThkSL01 and is
  91  * implemented as follows:
  92  * On entry, the EBX register is set up to contain a flat pointer to the
  93  * 16-bit stack such that EBX+22 points to the first argument.
  94  * Then, the entry point is called, while EBP is set up to point
  95  * to the return address (on the 32-bit stack).
  96  * The called function returns with CX set to the number of bytes
  97  * to be popped of the caller's stack.
  98  *
  99  * Stack layout upon entry to the core routine (STACK16FRAME):
 100  *  ...           ...
 101  * (sp+24) word   first 16-bit arg
 102  * (sp+22) word   cs
 103  * (sp+20) word   ip
 104  * (sp+18) word   bp
 105  * (sp+14) long   32-bit entry point (reused for Win16 mutex recursion count)
 106  * (sp+12) word   ip of actual entry point (necessary for relay debugging)
 107  * (sp+8)  long   relay (argument conversion) function entry point
 108  * (sp+4)  long   cs of 16-bit entry point
 109  * (sp)    long   ip of 16-bit entry point
 110  *
 111  * Added on the stack:
 112  * (sp-2)  word   saved gs
 113  * (sp-4)  word   saved fs
 114  * (sp-6)  word   saved es
 115  * (sp-8)  word   saved ds
 116  * (sp-12) long   saved ebp
 117  * (sp-16) long   saved ecx
 118  * (sp-20) long   saved edx
 119  * (sp-24) long   saved previous stack
 120  */
 121 static void BuildCallFrom16Core( FILE *outfile, int reg_func, int thunk, int short_ret )
 122 {
 123     char *name = thunk? "thunk" : reg_func? "regs" : short_ret? "word" : "long";
 124
 125     /* Function header */
 126     if (thunk) function_header( outfile, "__wine_call_from_16_thunk" );
 127     else if (reg_func) function_header( outfile, "__wine_call_from_16_regs" );
 128     else if (short_ret) function_header( outfile, "__wine_call_from_16_word" );
 129     else function_header( outfile, "__wine_call_from_16_long" );
 130
 131     /* Create STACK16FRAME (except STACK32FRAME link) */
 132     fprintf( outfile, "\tpushw %%gs\n" );
 133     fprintf( outfile, "\tpushw %%fs\n" );
 134     fprintf( outfile, "\tpushw %%es\n" );
 135     fprintf( outfile, "\tpushw %%ds\n" );
 136     fprintf( outfile, "\tpushl %%ebp\n" );
 137     fprintf( outfile, "\tpushl %%ecx\n" );
 138     fprintf( outfile, "\tpushl %%edx\n" );
 139
 140     /* Save original EFlags register */
 141     fprintf( outfile, "\tpushfl\n" );
 142
 143     if ( UsePIC )
 144     {
 145         /* Get Global Offset Table into %ecx */
 146         fprintf( outfile, "\tcall .L__wine_call_from_16_%s.getgot1\n", name );
 147         fprintf( outfile, ".L__wine_call_from_16_%s.getgot1:\n", name );
 148         fprintf( outfile, "\tpopl %%ecx\n" );
 149         fprintf( outfile, "\taddl $_GLOBAL_OFFSET_TABLE_+[.-.L__wine_call_from_16_%s.getgot1], %%ecx\n", name );
 150     }
 151
 152     if (UsePIC)
 153     {
 154         fprintf( outfile, "\t.byte 0x2e\n\tmovl " __ASM_NAME("CallTo16_DataSelector@GOT") "(%%ecx), %%edx\n" );
 155         fprintf( outfile, "\t.byte 0x2e\n\tmovl (%%edx), %%edx\n" );
 156     }
 157     else
 158         fprintf( outfile, "\t.byte 0x2e\n\tmovl " __ASM_NAME("CallTo16_DataSelector") ",%%edx\n" );
 159
 160     /* Load 32-bit segment registers */
 161 #ifdef __svr4__
 162     fprintf( outfile, "\tdata16\n");
 163 #endif
 164     fprintf( outfile, "\tmovw %%dx, %%ds\n" );
 165 #ifdef __svr4__
 166     fprintf( outfile, "\tdata16\n");
 167 #endif
 168     fprintf( outfile, "\tmovw %%dx, %%es\n" );
 169
 170     if ( UsePIC )
 171     {
 172         fprintf( outfile, "\tmovl " __ASM_NAME("SYSLEVEL_Win16CurrentTeb@GOT") "(%%ecx), %%edx\n" );
 173         fprintf( outfile, "\tmovw (%%edx), %%fs\n" );
 174     }
 175     else
 176         fprintf( outfile, "\tmovw " __ASM_NAME("SYSLEVEL_Win16CurrentTeb") ", %%fs\n" );
 177
 178     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d),%%gs\n", STRUCTOFFSET(TEB,gs_sel) );
 179
 180     /* Get address of wine_ldt_copy array into %ecx */
 181     if ( UsePIC )
 182         fprintf( outfile, "\tmovl " __ASM_NAME("wine_ldt_copy@GOT") "(%%ecx), %%ecx\n" );
 183     else
 184         fprintf( outfile, "\tmovl $" __ASM_NAME("wine_ldt_copy") ", %%ecx\n" );
 185
 186     /* Translate STACK16FRAME base to flat offset in %edx */
 187     fprintf( outfile, "\tmovw %%ss, %%dx\n" );
 188     fprintf( outfile, "\tandl $0xfff8, %%edx\n" );
 189     fprintf( outfile, "\tshrl $1, %%edx\n" );
 190     fprintf( outfile, "\tmovl (%%ecx,%%edx), %%edx\n" );
 191     fprintf( outfile, "\tmovzwl %%sp, %%ebp\n" );
 192     fprintf( outfile, "\tleal (%%ebp,%%edx), %%edx\n" );
 193
 194     /* Get saved flags into %ecx */
 195     fprintf( outfile, "\tpopl %%ecx\n" );
 196
 197     /* Get the 32-bit stack pointer from the TEB and complete STACK16FRAME */
 198     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d), %%ebp\n", STACKOFFSET );
 199     fprintf( outfile, "\tpushl %%ebp\n" );
 200
 201     /* Switch stacks */
 202 #ifdef __svr4__
 203     fprintf( outfile,"\tdata16\n");
 204 #endif
 205     fprintf( outfile, "\t.byte 0x64\n\tmovw %%ss, (%d)\n", STACKOFFSET + 2 );
 206     fprintf( outfile, "\t.byte 0x64\n\tmovw %%sp, (%d)\n", STACKOFFSET );
 207     fprintf( outfile, "\tpushl %%ds\n" );
 208     fprintf( outfile, "\tpopl %%ss\n" );
 209     fprintf( outfile, "\tmovl %%ebp, %%esp\n" );
 210     fprintf( outfile, "\taddl $%d, %%ebp\n", STRUCTOFFSET(STACK32FRAME, ebp) );
 211
 212
 213     /* At this point:
 214        STACK16FRAME is completely set up
 215        DS, ES, SS: flat data segment
 216        FS: current TEB
 217        ESP: points to last STACK32FRAME
 218        EBP: points to ebp member of last STACK32FRAME
 219        EDX: points to current STACK16FRAME
 220        ECX: contains saved flags
 221        all other registers: unchanged */
 222
 223     /* Special case: C16ThkSL stub */
 224     if ( thunk )
 225     {
 226         /* Set up registers as expected and call thunk */
 227         fprintf( outfile, "\tleal %d(%%edx), %%ebx\n", sizeof(STACK16FRAME)-22 );
 228         fprintf( outfile, "\tleal -4(%%esp), %%ebp\n" );
 229
 230         fprintf( outfile, "\tcall *%d(%%edx)\n", STACK16OFFSET(entry_point) );
 231
 232         /* Switch stack back */
 233         fprintf( outfile, "\t.byte 0x64\n\tmovw (%d), %%ss\n", STACKOFFSET+2 );
 234         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%esp\n", STACKOFFSET );
 235         fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 236
 237         /* Restore registers and return directly to caller */
 238         fprintf( outfile, "\taddl $8, %%esp\n" );
 239         fprintf( outfile, "\tpopl %%ebp\n" );
 240         fprintf( outfile, "\tpopw %%ds\n" );
 241         fprintf( outfile, "\tpopw %%es\n" );
 242         fprintf( outfile, "\tpopw %%fs\n" );
 243         fprintf( outfile, "\tpopw %%gs\n" );
 244         fprintf( outfile, "\taddl $20, %%esp\n" );
 245
 246         fprintf( outfile, "\txorb %%ch, %%ch\n" );
 247         fprintf( outfile, "\tpopl %%ebx\n" );
 248         fprintf( outfile, "\taddw %%cx, %%sp\n" );
 249         fprintf( outfile, "\tpush %%ebx\n" );
 250
 251         fprintf( outfile, "\t.byte 0x66\n" );
 252         fprintf( outfile, "\tlret\n" );
 253
 254         return;
 255     }
 256
 257
 258     /* Build register CONTEXT */
 259     if ( reg_func )
 260     {
 261         fprintf( outfile, "\tsubl $%d, %%esp\n", sizeof(CONTEXT86) );
 262
 263         fprintf( outfile, "\tmovl %%ecx, %d(%%esp)\n", CONTEXTOFFSET(EFlags) );
 264
 265         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Eax) );
 266         fprintf( outfile, "\tmovl %%ebx, %d(%%esp)\n", CONTEXTOFFSET(Ebx) );
 267         fprintf( outfile, "\tmovl %%esi, %d(%%esp)\n", CONTEXTOFFSET(Esi) );
 268         fprintf( outfile, "\tmovl %%edi, %d(%%esp)\n", CONTEXTOFFSET(Edi) );
 269
 270         fprintf( outfile, "\tmovl %d(%%edx), %%eax\n", STACK16OFFSET(ebp) );
 271         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Ebp) );
 272         fprintf( outfile, "\tmovl %d(%%edx), %%eax\n", STACK16OFFSET(ecx) );
 273         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Ecx) );
 274         fprintf( outfile, "\tmovl %d(%%edx), %%eax\n", STACK16OFFSET(edx) );
 275         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Edx) );
 276
 277         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(ds) );
 278         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegDs) );
 279         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(es) );
 280         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegEs) );
 281         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(fs) );
 282         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegFs) );
 283         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(gs) );
 284         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegGs) );
 285
 286         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(cs) );
 287         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegCs) );
 288         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(ip) );
 289         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Eip) );
 290
 291         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%eax\n", STACKOFFSET+2 );
 292         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegSs) );
 293         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%eax\n", STACKOFFSET );
 294         fprintf( outfile, "\taddl $%d, %%eax\n", STACK16OFFSET(ip) );
 295         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Esp) );
 296 #if 0
 297         fprintf( outfile, "\tfsave %d(%%esp)\n", CONTEXTOFFSET(FloatSave) );
 298 #endif
 299
 300         /* Push address of CONTEXT86 structure -- popped by the relay routine */
 301         fprintf( outfile, "\tpushl %%esp\n" );
 302     }
 303
 304
 305     /* Print debug info before call */
 306     if ( debugging )
 307     {
 308         if ( UsePIC )
 309         {
 310             fprintf( outfile, "\tpushl %%ebx\n" );
 311
 312             /* Get Global Offset Table into %ebx (for PLT call) */
 313             fprintf( outfile, "\tcall .L__wine_call_from_16_%s.getgot2\n", name );
 314             fprintf( outfile, ".L__wine_call_from_16_%s.getgot2:\n", name );
 315             fprintf( outfile, "\tpopl %%ebx\n" );
 316             fprintf( outfile, "\taddl $_GLOBAL_OFFSET_TABLE_+[.-.L__wine_call_from_16_%s.getgot2], %%ebx\n", name );
 317         }
 318
 319         fprintf( outfile, "\tpushl %%edx\n" );
 320         if ( reg_func )
 321             fprintf( outfile, "\tleal -%d(%%ebp), %%eax\n\tpushl %%eax\n",
 322                               sizeof(CONTEXT) + STRUCTOFFSET(STACK32FRAME, ebp) );
 323         else
 324             fprintf( outfile, "\tpushl $0\n" );
 325
 326         if ( UsePIC )
 327             fprintf( outfile, "\tcall " __ASM_NAME("RELAY_DebugCallFrom16@PLT") "\n ");
 328         else
 329             fprintf( outfile, "\tcall " __ASM_NAME("RELAY_DebugCallFrom16") "\n ");
 330
 331         fprintf( outfile, "\tpopl %%edx\n" );
 332         fprintf( outfile, "\tpopl %%edx\n" );
 333
 334         if ( UsePIC )
 335             fprintf( outfile, "\tpopl %%ebx\n" );
 336     }
 337
 338     /* Call relay routine (which will call the API entry point) */
 339     fprintf( outfile, "\tleal %d(%%edx), %%eax\n", sizeof(STACK16FRAME) );
 340     fprintf( outfile, "\tpushl %%eax\n" );
 341     fprintf( outfile, "\tpushl %d(%%edx)\n", STACK16OFFSET(entry_point) );
 342     fprintf( outfile, "\tcall *%d(%%edx)\n", STACK16OFFSET(relay) );
 343
 344     /* Print debug info after call */
 345     if ( debugging )
 346     {
 347         if ( UsePIC )
 348         {
 349             fprintf( outfile, "\tpushl %%ebx\n" );
 350
 351             /* Get Global Offset Table into %ebx (for PLT call) */
 352             fprintf( outfile, "\tcall .L__wine_call_from_16_%s.getgot3\n", name );
 353             fprintf( outfile, ".L__wine_call_from_16_%s.getgot3:\n", name );
 354             fprintf( outfile, "\tpopl %%ebx\n" );
 355             fprintf( outfile, "\taddl $_GLOBAL_OFFSET_TABLE_+[.-.L__wine_call_from_16_%s.getgot3], %%ebx\n", name );
 356         }
 357
 358         fprintf( outfile, "\tpushl %%eax\n" );
 359         if ( reg_func )
 360             fprintf( outfile, "\tleal -%d(%%ebp), %%eax\n\tpushl %%eax\n",
 361                               sizeof(CONTEXT) + STRUCTOFFSET(STACK32FRAME, ebp) );
 362         else
 363             fprintf( outfile, "\tpushl $0\n" );
 364
 365         if ( UsePIC )
 366             fprintf( outfile, "\tcall " __ASM_NAME("RELAY_DebugCallFrom16Ret@PLT") "\n ");
 367         else
 368             fprintf( outfile, "\tcall " __ASM_NAME("RELAY_DebugCallFrom16Ret") "\n ");
 369
 370         fprintf( outfile, "\tpopl %%eax\n" );
 371         fprintf( outfile, "\tpopl %%eax\n" );
 372
 373         if ( UsePIC )
 374             fprintf( outfile, "\tpopl %%ebx\n" );
 375     }
 376
 377
 378     if ( reg_func )
 379     {
 380         fprintf( outfile, "\tmovl %%esp, %%ebx\n" );
 381
 382         /* Switch stack back */
 383         fprintf( outfile, "\t.byte 0x64\n\tmovw (%d), %%ss\n", STACKOFFSET+2 );
 384         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%esp\n", STACKOFFSET );
 385         fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 386
 387         /* Get return address to CallFrom16 stub */
 388         fprintf( outfile, "\taddw $%d, %%sp\n", STACK16OFFSET(callfrom_ip)-4 );
 389         fprintf( outfile, "\tpopl %%eax\n" );
 390         fprintf( outfile, "\tpopl %%edx\n" );
 391
 392         /* Restore all registers from CONTEXT */
 393         fprintf( outfile, "\tmovw %d(%%ebx), %%ss\n", CONTEXTOFFSET(SegSs) );
 394         fprintf( outfile, "\tmovl %d(%%ebx), %%esp\n", CONTEXTOFFSET(Esp) );
 395         fprintf( outfile, "\taddl $4, %%esp\n" );  /* room for final return address */
 396
 397         fprintf( outfile, "\tpushw %d(%%ebx)\n", CONTEXTOFFSET(SegCs) );
 398         fprintf( outfile, "\tpushw %d(%%ebx)\n", CONTEXTOFFSET(Eip) );
 399         fprintf( outfile, "\tpushl %%edx\n" );
 400         fprintf( outfile, "\tpushl %%eax\n" );
 401         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(EFlags) );
 402         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(SegDs) );
 403
 404         fprintf( outfile, "\tmovw %d(%%ebx), %%es\n", CONTEXTOFFSET(SegEs) );
 405         fprintf( outfile, "\tmovw %d(%%ebx), %%fs\n", CONTEXTOFFSET(SegFs) );
 406         fprintf( outfile, "\tmovw %d(%%ebx), %%gs\n", CONTEXTOFFSET(SegGs) );
 407
 408         fprintf( outfile, "\tmovl %d(%%ebx), %%ebp\n", CONTEXTOFFSET(Ebp) );
 409         fprintf( outfile, "\tmovl %d(%%ebx), %%esi\n", CONTEXTOFFSET(Esi) );
 410         fprintf( outfile, "\tmovl %d(%%ebx), %%edi\n", CONTEXTOFFSET(Edi) );
 411         fprintf( outfile, "\tmovl %d(%%ebx), %%eax\n", CONTEXTOFFSET(Eax) );
 412         fprintf( outfile, "\tmovl %d(%%ebx), %%edx\n", CONTEXTOFFSET(Edx) );
 413         fprintf( outfile, "\tmovl %d(%%ebx), %%ecx\n", CONTEXTOFFSET(Ecx) );
 414         fprintf( outfile, "\tmovl %d(%%ebx), %%ebx\n", CONTEXTOFFSET(Ebx) );
 415
 416         fprintf( outfile, "\tpopl %%ds\n" );
 417         fprintf( outfile, "\tpopfl\n" );
 418         fprintf( outfile, "\tlret\n" );
 419     }
 420     else
 421     {
 422         /* Switch stack back */
 423         fprintf( outfile, "\t.byte 0x64\n\tmovw (%d), %%ss\n", STACKOFFSET+2 );
 424         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%esp\n", STACKOFFSET );
 425         fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 426
 427         /* Restore registers */
 428         fprintf( outfile, "\tpopl %%edx\n" );
 429         fprintf( outfile, "\tpopl %%ecx\n" );
 430         fprintf( outfile, "\tpopl %%ebp\n" );
 431         fprintf( outfile, "\tpopw %%ds\n" );
 432         fprintf( outfile, "\tpopw %%es\n" );
 433         fprintf( outfile, "\tpopw %%fs\n" );
 434         fprintf( outfile, "\tpopw %%gs\n" );
 435
 436         /* Prepare return value and set flags accordingly */
 437         if ( !short_ret )
 438             fprintf( outfile, "\tshldl $16, %%eax, %%edx\n" );
 439         fprintf( outfile, "\torl %%eax, %%eax\n" );
 440
 441         /* Return to return stub which will return to caller */
 442         fprintf( outfile, "\tlret $12\n" );
 443     }
 444 }
 445
 446
 447 /*******************************************************************
 448  *         BuildCallTo16Core
 449  *
 450  * This routine builds the core routines used in 32->16 thunks:
 451  *
 452  *   extern LONG WINAPI wine_call_to_16( SEGPTR target, int nb_args );
 453  *   extern void WINAPI wine_call_to_16_regs_short( const CONTEXT86 *context, int nb_args );
 454  *   extern void WINAPI wine_call_to_16_regs_long ( const CONTEXT86 *context, int nb_args );
 455  *
 456  * These routines can be called directly from 32-bit code.
 457  *
 458  * All routines expect that the 16-bit stack contents (arguments) were
 459  * already set up by the caller; nb_args must contain the number of bytes
 460  * to be conserved.  The 16-bit SS:SP will be set accordinly.
 461  *
 462  * All other registers are either taken from the CONTEXT86 structure
 463  * or else set to default values.  The target routine address is either
 464  * given directly or taken from the CONTEXT86.
 465  *
 466  * If you want to call a 16-bit routine taking only standard argument types
 467  * (WORD and LONG), you can also have an appropriate argument conversion
 468  * stub automatically generated (see BuildCallTo16); you'd then call this
 469  * stub, which in turn would prepare the 16-bit stack and call the appropiate
 470  * core routine.
 471  *
 472  */
 473 static void BuildCallTo16Core( FILE *outfile, int reg_func )
 474 {
 475     const char *name = reg_func == 2 ? "wine_call_to_16_regs_long" :
 476                        reg_func == 1 ? "wine_call_to_16_regs_short" :
 477                        "wine_call_to_16";
 478
 479     /* Function header */
 480     function_header( outfile, name );
 481
 482     /* Function entry sequence */
 483     fprintf( outfile, "\tpushl %%ebp\n" );
 484     fprintf( outfile, "\tmovl %%esp, %%ebp\n" );
 485
 486     /* Save the 32-bit registers */
 487     fprintf( outfile, "\tpushl %%ebx\n" );
 488     fprintf( outfile, "\tpushl %%esi\n" );
 489     fprintf( outfile, "\tpushl %%edi\n" );
 490     fprintf( outfile, "\t.byte 0x64\n\tmovl %%gs,(%d)\n", STRUCTOFFSET(TEB,gs_sel) );
 491
 492     if ( UsePIC )
 493     {
 494         /* Get Global Offset Table into %ebx */
 495         fprintf( outfile, "\tcall .L%s.getgot1\n", name );
 496         fprintf( outfile, ".L%s.getgot1:\n", name );
 497         fprintf( outfile, "\tpopl %%ebx\n" );
 498         fprintf( outfile, "\taddl $_GLOBAL_OFFSET_TABLE_+[.-.L%s.getgot1], %%ebx\n", name );
 499     }
 500
 501     /* Setup exception frame */
 502     fprintf( outfile, "\t.byte 0x64\n\tpushl (%d)\n", STACKOFFSET );
 503     if (UsePIC)
 504         fprintf( outfile, "\tpushl " __ASM_NAME("__wine_callto16_handler@GOT") "(%%ebx)\n" );
 505     else
 506         fprintf( outfile, "\tpushl $" __ASM_NAME("__wine_callto16_handler") "\n" );
 507     fprintf( outfile, "\t.byte 0x64\n\tpushl (%d)\n", STRUCTOFFSET(TEB,except) );
 508     fprintf( outfile, "\t.byte 0x64\n\tmovl %%esp,(%d)\n", STRUCTOFFSET(TEB,except) );
 509
 510     /* Get return address */
 511     if ( UsePIC )
 512     {
 513         fprintf( outfile, "\tmovl " __ASM_NAME("CallTo16_RetAddr@GOT") "(%%ebx), %%ecx\n" );
 514         fprintf( outfile, "\tmovl (%%ecx), %%ecx\n" );
 515     }
 516     else
 517         fprintf( outfile, "\tmovl " __ASM_NAME("CallTo16_RetAddr") ", %%ecx\n" );
 518
 519     /* Call the actual CallTo16 routine (simulate a lcall) */
 520     fprintf( outfile, "\tpushl %%cs\n" );
 521     fprintf( outfile, "\tcall .L%s\n", name );
 522
 523     /* Remove exception frame */
 524     fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STRUCTOFFSET(TEB,except) );
 525     fprintf( outfile, "\taddl $4, %%esp\n" );
 526     fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 527
 528     if ( !reg_func )
 529     {
 530         /* Convert return value */
 531         fprintf( outfile, "\tshll $16,%%edx\n" );
 532         fprintf( outfile, "\tmovw %%ax,%%dx\n" );
 533         fprintf( outfile, "\tmovl %%edx,%%eax\n" );
 534     }
 535     else
 536     {
 537         /*
 538          * Modify CONTEXT86 structure to contain new values
 539          *
 540          * NOTE:  We restore only EAX, EBX, EDX, EDX, EBP, and ESP.
 541          *        The segment registers as well as ESI and EDI should
 542          *        not be modified by a well-behaved 16-bit routine in
 543          *        any case.  [If necessary, we could restore them as well,
 544          *        at the cost of a somewhat less efficient return path.]
 545          */
 546
 547         fprintf( outfile, "\tmovl %d(%%esp), %%edi\n", STACK32OFFSET(target) - STACK32OFFSET(edi));
 548                 /* everything above edi has been popped already */
 549
 550         fprintf( outfile, "\tmovl %%eax, %d(%%edi)\n", CONTEXTOFFSET(Eax) );
 551         fprintf( outfile, "\tmovl %%ebx, %d(%%edi)\n", CONTEXTOFFSET(Ebx) );
 552         fprintf( outfile, "\tmovl %%ecx, %d(%%edi)\n", CONTEXTOFFSET(Ecx) );
 553         fprintf( outfile, "\tmovl %%edx, %d(%%edi)\n", CONTEXTOFFSET(Edx) );
 554         fprintf( outfile, "\tmovl %%ebp, %d(%%edi)\n", CONTEXTOFFSET(Ebp) );
 555         fprintf( outfile, "\tmovl %%esi, %d(%%edi)\n", CONTEXTOFFSET(Esp) );
 556                  /* The return glue code saved %esp into %esi */
 557     }
 558
 559     /* Restore the 32-bit registers */
 560     fprintf( outfile, "\tpopl %%edi\n" );
 561     fprintf( outfile, "\tpopl %%esi\n" );
 562     fprintf( outfile, "\tpopl %%ebx\n" );
 563
 564     /* Function exit sequence */
 565     fprintf( outfile, "\tpopl %%ebp\n" );
 566     fprintf( outfile, "\tret $8\n" );
 567
 568
 569     /* Start of the actual CallTo16 routine */
 570
 571     fprintf( outfile, ".L%s:\n", name );
 572
 573     /* Switch to the 16-bit stack */
 574     fprintf( outfile, "\tmovl %%esp,%%edx\n" );
 575 #ifdef __svr4__
 576     fprintf( outfile,"\tdata16\n");
 577 #endif
 578     fprintf( outfile, "\t.byte 0x64\n\tmovw (%d),%%ss\n", STACKOFFSET + 2);
 579     fprintf( outfile, "\t.byte 0x64\n\tmovw (%d),%%sp\n", STACKOFFSET );
 580     fprintf( outfile, "\t.byte 0x64\n\tmovl %%edx,(%d)\n", STACKOFFSET );
 581
 582     /* Make %bp point to the previous stackframe (built by CallFrom16) */
 583     fprintf( outfile, "\tmovzwl %%sp,%%ebp\n" );
 584     fprintf( outfile, "\tleal %d(%%ebp),%%ebp\n", STACK16OFFSET(bp) );
 585
 586     /* Add the specified offset to the new sp */
 587     fprintf( outfile, "\tsubw %d(%%edx), %%sp\n", STACK32OFFSET(nb_args) );
 588
 589     /* Push the return address
 590      * With sreg suffix, we push 16:16 address (normal lret)
 591      * With lreg suffix, we push 16:32 address (0x66 lret, for KERNEL32_45)
 592      */
 593     if (reg_func != 2)
 594         fprintf( outfile, "\tpushl %%ecx\n" );
 595     else
 596     {
 597         fprintf( outfile, "\tshldl $16, %%ecx, %%eax\n" );
 598         fprintf( outfile, "\tpushw $0\n" );
 599         fprintf( outfile, "\tpushw %%ax\n" );
 600         fprintf( outfile, "\tpushw $0\n" );
 601         fprintf( outfile, "\tpushw %%cx\n" );
 602     }
 603
 604     if (reg_func)
 605     {
 606         /* Push the called routine address */
 607         fprintf( outfile, "\tmovl %d(%%edx),%%edx\n", STACK32OFFSET(target) );
 608         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(SegCs) );
 609         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(Eip) );
 610
 611         /* Get the registers */
 612         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(SegDs) );
 613         fprintf( outfile, "\tmovl %d(%%edx),%%eax\n", CONTEXTOFFSET(SegEs) );
 614         fprintf( outfile, "\tmovw %%ax,%%es\n" );
 615         fprintf( outfile, "\tmovl %d(%%edx),%%eax\n", CONTEXTOFFSET(SegFs) );
 616         fprintf( outfile, "\tmovw %%ax,%%fs\n" );
 617         fprintf( outfile, "\tmovl %d(%%edx),%%eax\n", CONTEXTOFFSET(SegGs) );
 618         fprintf( outfile, "\tmovw %%ax,%%gs\n" );
 619         fprintf( outfile, "\tmovl %d(%%edx),%%ebp\n", CONTEXTOFFSET(Ebp) );
 620         fprintf( outfile, "\tmovl %d(%%edx),%%esi\n", CONTEXTOFFSET(Esi) );
 621         fprintf( outfile, "\tmovl %d(%%edx),%%edi\n", CONTEXTOFFSET(Edi) );
 622         fprintf( outfile, "\tmovl %d(%%edx),%%eax\n", CONTEXTOFFSET(Eax) );
 623         fprintf( outfile, "\tmovl %d(%%edx),%%ebx\n", CONTEXTOFFSET(Ebx) );
 624         fprintf( outfile, "\tmovl %d(%%edx),%%ecx\n", CONTEXTOFFSET(Ecx) );
 625         fprintf( outfile, "\tmovl %d(%%edx),%%edx\n", CONTEXTOFFSET(Edx) );
 626
 627         /* Get the 16-bit ds */
 628         fprintf( outfile, "\tpopw %%ds\n" );
 629     }
 630     else  /* not a register function */
 631     {
 632         /* Push the called routine address */
 633         fprintf( outfile, "\tpushl %d(%%edx)\n", STACK32OFFSET(target) );
 634
 635         /* Set %fs and %gs to the value saved by the last CallFrom16 */
 636         fprintf( outfile, "\tmovw %d(%%ebp),%%ax\n", STACK16OFFSET(fs)-STACK16OFFSET(bp) );
 637         fprintf( outfile, "\tmovw %%ax,%%fs\n" );
 638         fprintf( outfile, "\tmovw %d(%%ebp),%%ax\n", STACK16OFFSET(gs)-STACK16OFFSET(bp) );
 639         fprintf( outfile, "\tmovw %%ax,%%gs\n" );
 640
 641         /* Set %ds and %es (and %ax just in case) equal to %ss */
 642         fprintf( outfile, "\tmovw %%ss,%%ax\n" );
 643         fprintf( outfile, "\tmovw %%ax,%%ds\n" );
 644         fprintf( outfile, "\tmovw %%ax,%%es\n" );
 645     }
 646
 647     /* Jump to the called routine */
 648     fprintf( outfile, "\t.byte 0x66\n" );
 649     fprintf( outfile, "\tlret\n" );
 650 }
 651
 652
 653 /*******************************************************************
 654  *         BuildRet16Func
 655  *
 656  * Build the return code for 16-bit callbacks
 657  */
 658 static void BuildRet16Func( FILE *outfile )
 659 {
 660     /*
 661      *  Note: This must reside in the .data section to allow
 662      *        run-time relocation of the SYSLEVEL_Win16CurrentTeb symbol
 663      */
 664
 665     function_header( outfile, "CallTo16_Ret" );
 666
 667     /* Save %esp into %esi */
 668     fprintf( outfile, "\tmovl %%esp,%%esi\n" );
 669
 670     /* Restore 32-bit segment registers */
 671
 672     fprintf( outfile, "\t.byte 0x2e\n\tmovl " __ASM_NAME("CallTo16_DataSelector") "-" __ASM_NAME("Call16_Ret_Start") ",%%edi\n" );
 673 #ifdef __svr4__
 674     fprintf( outfile, "\tdata16\n");
 675 #endif
 676     fprintf( outfile, "\tmovw %%di,%%ds\n" );
 677 #ifdef __svr4__
 678     fprintf( outfile, "\tdata16\n");
 679 #endif
 680     fprintf( outfile, "\tmovw %%di,%%es\n" );
 681
 682     fprintf( outfile, "\tmovw " __ASM_NAME("SYSLEVEL_Win16CurrentTeb") ",%%fs\n" );
 683
 684     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d),%%gs\n", STRUCTOFFSET(TEB,gs_sel) );
 685
 686     /* Restore the 32-bit stack */
 687
 688 #ifdef __svr4__
 689     fprintf( outfile, "\tdata16\n");
 690 #endif
 691     fprintf( outfile, "\tmovw %%di,%%ss\n" );
 692     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d),%%esp\n", STACKOFFSET );
 693
 694     /* Return to caller */
 695
 696     fprintf( outfile, "\tlret\n" );
 697
 698     /* Declare the return address and data selector variables */
 699
 700     fprintf( outfile, "\n\t.align %d\n", get_alignment(4) );
 701     fprintf( outfile, "\t.globl " __ASM_NAME("CallTo16_DataSelector") "\n" );
 702     fprintf( outfile, __ASM_NAME("CallTo16_DataSelector") ":\t.long 0\n" );
 703     fprintf( outfile, "\t.globl " __ASM_NAME("CallTo16_RetAddr") "\n" );
 704     fprintf( outfile, __ASM_NAME("CallTo16_RetAddr") ":\t.long 0\n" );
 705 }
 706
 707
 708 /*******************************************************************
 709  *         BuildCallTo32CBClient
 710  *
 711  * Call a CBClient relay stub from 32-bit code (KERNEL.620).
 712  *
 713  * Since the relay stub is itself 32-bit, this should not be a problem;
 714  * unfortunately, the relay stubs are expected to switch back to a
 715  * 16-bit stack (and 16-bit code) after completion :-(
 716  *
 717  * This would conflict with our 16- vs. 32-bit stack handling, so
 718  * we simply switch *back* to our 32-bit stack before returning to
 719  * the caller ...
 720  *
 721  * The CBClient relay stub expects to be called with the following
 722  * 16-bit stack layout, and with ebp and ebx pointing into the 16-bit
 723  * stack at the designated places:
 724  *
 725  *    ...
 726  *  (ebp+14) original arguments to the callback routine
 727  *  (ebp+10) far return address to original caller
 728  *  (ebp+6)  Thunklet target address
 729  *  (ebp+2)  Thunklet relay ID code
 730  *  (ebp)    BP (saved by CBClientGlueSL)
 731  *  (ebp-2)  SI (saved by CBClientGlueSL)
 732  *  (ebp-4)  DI (saved by CBClientGlueSL)
 733  *  (ebp-6)  DS (saved by CBClientGlueSL)
 734  *
 735  *   ...     buffer space used by the 16-bit side glue for temp copies
 736  *
 737  *  (ebx+4)  far return address to 16-bit side glue code
 738  *  (ebx)    saved 16-bit ss:sp (pointing to ebx+4)
 739  *
 740  * The 32-bit side glue code accesses both the original arguments (via ebp)
 741  * and the temporary copies prepared by the 16-bit side glue (via ebx).
 742  * After completion, the stub will load ss:sp from the buffer at ebx
 743  * and perform a far return to 16-bit code.
 744  *
 745  * To trick the relay stub into returning to us, we replace the 16-bit
 746  * return address to the glue code by a cs:ip pair pointing to our
 747  * return entry point (the original return address is saved first).
 748  * Our return stub thus called will then reload the 32-bit ss:esp and
 749  * return to 32-bit code (by using and ss:esp value that we have also
 750  * pushed onto the 16-bit stack before and a cs:eip values found at
 751  * that position on the 32-bit stack).  The ss:esp to be restored is
 752  * found relative to the 16-bit stack pointer at:
 753  *
 754  *  (ebx-4)   ss  (flat)
 755  *  (ebx-8)   sp  (32-bit stack pointer)
 756  *
 757  * The second variant of this routine, CALL32_CBClientEx, which is used
 758  * to implement KERNEL.621, has to cope with yet another problem: Here,
 759  * the 32-bit side directly returns to the caller of the CBClient thunklet,
 760  * restoring registers saved by CBClientGlueSL and cleaning up the stack.
 761  * As we have to return to our 32-bit code first, we have to adapt the
 762  * layout of our temporary area so as to include values for the registers
 763  * that are to be restored, and later (in the implementation of KERNEL.621)
 764  * we *really* restore them. The return stub restores DS, DI, SI, and BP
 765  * from the stack, skips the next 8 bytes (CBClient relay code / target),
 766  * and then performs a lret NN, where NN is the number of arguments to be
 767  * removed. Thus, we prepare our temporary area as follows:
 768  *
 769  *     (ebx+22) 16-bit cs  (this segment)
 770  *     (ebx+20) 16-bit ip  ('16-bit' return entry point)
 771  *     (ebx+16) 32-bit ss  (flat)
 772  *     (ebx+12) 32-bit sp  (32-bit stack pointer)
 773  *     (ebx+10) 16-bit bp  (points to ebx+24)
 774  *     (ebx+8)  16-bit si  (ignored)
 775  *     (ebx+6)  16-bit di  (ignored)
 776  *     (ebx+4)  16-bit ds  (we actually use the flat DS here)
 777  *     (ebx+2)  16-bit ss  (16-bit stack segment)
 778  *     (ebx+0)  16-bit sp  (points to ebx+4)
 779  *
 780  * Note that we ensure that DS is not changed and remains the flat segment,
 781  * and the 32-bit stack pointer our own return stub needs fits just
 782  * perfectly into the 8 bytes that are skipped by the Windows stub.
 783  * One problem is that we have to determine the number of removed arguments,
 784  * as these have to be really removed in KERNEL.621. Thus, the BP value
 785  * that we place in the temporary area to be restored, contains the value
 786  * that SP would have if no arguments were removed. By comparing the actual
 787  * value of SP with this value in our return stub we can compute the number
 788  * of removed arguments. This is then returned to KERNEL.621.
 789  *
 790  * The stack layout of this function:
 791  * (ebp+20)  nArgs     pointer to variable receiving nr. of args (Ex only)
 792  * (ebp+16)  esi       pointer to caller's esi value
 793  * (ebp+12)  arg       ebp value to be set for relay stub
 794  * (ebp+8)   func      CBClient relay stub address
 795  * (ebp+4)   ret addr
 796  * (ebp)     ebp
 797  */
 798 static void BuildCallTo32CBClient( FILE *outfile, BOOL isEx )
 799 {
 800     char *name = isEx? "CBClientEx" : "CBClient";
 801     int size = isEx? 24 : 12;
 802
 803     /* Function header */
 804
 805     fprintf( outfile, "\n\t.align %d\n", get_alignment(4) );
 806     if (use_stabs)
 807         fprintf( outfile, ".stabs \"CALL32_%s:F1\",36,0,0," __ASM_NAME("CALL32_%s") "\n",
 808                  name, name );
 809     fprintf( outfile, "\t.globl " __ASM_NAME("CALL32_%s") "\n", name );
 810     fprintf( outfile, __ASM_NAME("CALL32_%s") ":\n", name );
 811
 812     /* Entry code */
 813
 814     fprintf( outfile, "\tpushl %%ebp\n" );
 815     fprintf( outfile, "\tmovl %%esp,%%ebp\n" );
 816     fprintf( outfile, "\tpushl %%edi\n" );
 817     fprintf( outfile, "\tpushl %%esi\n" );
 818     fprintf( outfile, "\tpushl %%ebx\n" );
 819
 820     /* Get the 16-bit stack */
 821
 822     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d),%%ebx\n", STACKOFFSET);
 823
 824     /* Convert it to a flat address */
 825
 826     fprintf( outfile, "\tshldl $16,%%ebx,%%eax\n" );
 827     fprintf( outfile, "\tandl $0xfff8,%%eax\n" );
 828     fprintf( outfile, "\tshrl $1,%%eax\n" );
 829     fprintf( outfile, "\tmovl " __ASM_NAME("wine_ldt_copy") "(%%eax),%%esi\n" );
 830     fprintf( outfile, "\tmovw %%bx,%%ax\n" );
 831     fprintf( outfile, "\taddl %%eax,%%esi\n" );
 832
 833     /* Allocate temporary area (simulate STACK16_PUSH) */
 834
 835     fprintf( outfile, "\tpushf\n" );
 836     fprintf( outfile, "\tcld\n" );
 837     fprintf( outfile, "\tleal -%d(%%esi), %%edi\n", size );
 838     fprintf( outfile, "\tmovl $%d, %%ecx\n", sizeof(STACK16FRAME) );
 839     fprintf( outfile, "\trep\n\tmovsb\n" );
 840     fprintf( outfile, "\tpopf\n" );
 841
 842     fprintf( outfile, "\t.byte 0x64\n\tsubw $%d,(%d)\n", size, STACKOFFSET );
 843
 844     fprintf( outfile, "\tpushl %%edi\n" );  /* remember address */
 845
 846     /* Set up temporary area */
 847
 848     if ( !isEx )
 849     {
 850         fprintf( outfile, "\tleal 4(%%edi), %%edi\n" );
 851
 852         fprintf( outfile, "\tleal -8(%%esp), %%eax\n" );
 853         fprintf( outfile, "\tmovl %%eax, -8(%%edi)\n" );    /* 32-bit sp */
 854
 855         fprintf( outfile, "\tmovw %%ss, %%ax\n" );
 856         fprintf( outfile, "\tandl $0x0000ffff, %%eax\n" );
 857         fprintf( outfile, "\tmovl %%eax, -4(%%edi)\n" );    /* 32-bit ss */
 858
 859         fprintf( outfile, "\taddl $%d, %%ebx\n", sizeof(STACK16FRAME)-size+4 + 4 );
 860         fprintf( outfile, "\tmovl %%ebx, 0(%%edi)\n" );    /* 16-bit ss:sp */
 861
 862         fprintf( outfile, "\tmovl " __ASM_NAME("CALL32_%s_RetAddr") ", %%eax\n", name );
 863         fprintf( outfile, "\tmovl %%eax, 4(%%edi)\n" );   /* overwrite return address */
 864     }
 865     else
 866     {
 867         fprintf( outfile, "\taddl $%d, %%ebx\n", sizeof(STACK16FRAME)-size+4 );
 868         fprintf( outfile, "\tmovl %%ebx, 0(%%edi)\n" );
 869
 870         fprintf( outfile, "\tmovw %%ds, %%ax\n" );
 871         fprintf( outfile, "\tmovw %%ax, 4(%%edi)\n" );
 872
 873         fprintf( outfile, "\taddl $20, %%ebx\n" );
 874         fprintf( outfile, "\tmovw %%bx, 10(%%edi)\n" );
 875
 876         fprintf( outfile, "\tleal -8(%%esp), %%eax\n" );
 877         fprintf( outfile, "\tmovl %%eax, 12(%%edi)\n" );
 878
 879         fprintf( outfile, "\tmovw %%ss, %%ax\n" );
 880         fprintf( outfile, "\tandl $0x0000ffff, %%eax\n" );
 881         fprintf( outfile, "\tmovl %%eax, 16(%%edi)\n" );
 882
 883         fprintf( outfile, "\tmovl " __ASM_NAME("CALL32_%s_RetAddr") ", %%eax\n", name );
 884         fprintf( outfile, "\tmovl %%eax, 20(%%edi)\n" );
 885     }
 886
 887     /* Set up registers and call CBClient relay stub (simulating a far call) */
 888
 889     fprintf( outfile, "\tmovl 16(%%ebp), %%esi\n" );
 890     fprintf( outfile, "\tmovl (%%esi), %%esi\n" );
 891
 892     fprintf( outfile, "\tmovl %%edi, %%ebx\n" );
 893     fprintf( outfile, "\tmovl 8(%%ebp), %%eax\n" );
 894     fprintf( outfile, "\tmovl 12(%%ebp), %%ebp\n" );
 895
 896     fprintf( outfile, "\tpushl %%cs\n" );
 897     fprintf( outfile, "\tcall *%%eax\n" );
 898
 899     /* Return new esi value to caller */
 900
 901     fprintf( outfile, "\tmovl 32(%%esp), %%edi\n" );
 902     fprintf( outfile, "\tmovl %%esi, (%%edi)\n" );
 903
 904     /* Cleanup temporary area (simulate STACK16_POP) */
 905
 906     fprintf( outfile, "\tpop %%esi\n" );
 907
 908     fprintf( outfile, "\tpushf\n" );
 909     fprintf( outfile, "\tstd\n" );
 910     fprintf( outfile, "\tdec %%esi\n" );
 911     fprintf( outfile, "\tleal %d(%%esi), %%edi\n", size );
 912     fprintf( outfile, "\tmovl $%d, %%ecx\n", sizeof(STACK16FRAME) );
 913     fprintf( outfile, "\trep\n\tmovsb\n" );
 914     fprintf( outfile, "\tpopf\n" );
 915
 916     fprintf( outfile, "\t.byte 0x64\n\taddw $%d,(%d)\n", size, STACKOFFSET );
 917
 918     /* Return argument size to caller */
 919     if ( isEx )
 920     {
 921         fprintf( outfile, "\tmovl 32(%%esp), %%ebx\n" );
 922         fprintf( outfile, "\tmovl %%ebp, (%%ebx)\n" );
 923     }
 924
 925     /* Restore registers and return */
 926
 927     fprintf( outfile, "\tpopl %%ebx\n" );
 928     fprintf( outfile, "\tpopl %%esi\n" );
 929     fprintf( outfile, "\tpopl %%edi\n" );
 930     fprintf( outfile, "\tpopl %%ebp\n" );
 931     fprintf( outfile, "\tret\n" );
 932 }
 933
 934 static void BuildCallTo32CBClientRet( FILE *outfile, BOOL isEx )
 935 {
 936     char *name = isEx? "CBClientEx" : "CBClient";
 937
 938     /* '16-bit' return stub */
 939
 940     fprintf( outfile, "\n\t.globl " __ASM_NAME("CALL32_%s_Ret") "\n", name );
 941     fprintf( outfile, __ASM_NAME("CALL32_%s_Ret") ":\n", name );
 942
 943     if ( !isEx )
 944     {
 945         fprintf( outfile, "\tmovzwl %%sp, %%ebx\n" );
 946         fprintf( outfile, "\tlssl %%ss:-16(%%ebx), %%esp\n" );
 947     }
 948     else
 949     {
 950         fprintf( outfile, "\tmovzwl %%bp, %%ebx\n" );
 951         fprintf( outfile, "\tsubw %%bp, %%sp\n" );
 952         fprintf( outfile, "\tmovzwl %%sp, %%ebp\n" );
 953         fprintf( outfile, "\tlssl %%ss:-12(%%ebx), %%esp\n" );
 954     }
 955     fprintf( outfile, "\tlret\n" );
 956
 957     /* Declare the return address variable */
 958
 959     fprintf( outfile, "\n\t.globl " __ASM_NAME("CALL32_%s_RetAddr") "\n", name );
 960     fprintf( outfile, __ASM_NAME("CALL32_%s_RetAddr") ":\t.long 0\n", name );
 961 }
 962
 963
 964 /*******************************************************************
 965  *         BuildCallFrom32Regs
 966  *
 967  * Build a 32-bit-to-Wine call-back function for a 'register' function.
 968  * 'args' is the number of dword arguments.
 969  *
 970  * Stack layout:
 971  *   ...
 972  * (ebp+12)  first arg
 973  * (ebp+8)   ret addr to user code
 974  * (ebp+4)   ret addr to relay code
 975  * (ebp+0)   saved ebp
 976  * (ebp-128) buffer area to allow stack frame manipulation
 977  * (ebp-332) CONTEXT86 struct
 978  * (ebp-336) CONTEXT86 *argument
 979  *  ....     other arguments copied from (ebp+12)
 980  *
 981  * The entry point routine is called with a CONTEXT* extra argument,
 982  * following the normal args. In this context structure, EIP_reg
 983  * contains the return address to user code, and ESP_reg the stack
 984  * pointer on return (with the return address and arguments already
 985  * removed).
 986  */
 987 static void BuildCallFrom32Regs( FILE *outfile )
 988 {
 989     static const int STACK_SPACE = 128 + sizeof(CONTEXT86);
 990
 991     /* Function header */
 992
 993     function_header( outfile, "__wine_call_from_32_regs" );
 994
 995     /* Allocate some buffer space on the stack */
 996
 997     fprintf( outfile, "\tpushl %%ebp\n" );
 998     fprintf( outfile, "\tmovl %%esp,%%ebp\n ");
 999     fprintf( outfile, "\tleal -%d(%%esp), %%esp\n", STACK_SPACE );
1000
1001     /* Build the context structure */
1002
1003     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Eax) - STACK_SPACE );
1004     fprintf( outfile, "\tpushfl\n" );
1005     fprintf( outfile, "\tpopl %%eax\n" );
1006     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(EFlags) - STACK_SPACE );
1007     fprintf( outfile, "\tmovl 0(%%ebp),%%eax\n" );
1008     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Ebp) - STACK_SPACE );
1009     fprintf( outfile, "\tmovl %%ebx,%d(%%ebp)\n", CONTEXTOFFSET(Ebx) - STACK_SPACE );
1010     fprintf( outfile, "\tmovl %%ecx,%d(%%ebp)\n", CONTEXTOFFSET(Ecx) - STACK_SPACE );
1011     fprintf( outfile, "\tmovl %%edx,%d(%%ebp)\n", CONTEXTOFFSET(Edx) - STACK_SPACE );
1012     fprintf( outfile, "\tmovl %%esi,%d(%%ebp)\n", CONTEXTOFFSET(Esi) - STACK_SPACE );
1013     fprintf( outfile, "\tmovl %%edi,%d(%%ebp)\n", CONTEXTOFFSET(Edi) - STACK_SPACE );
1014
1015     fprintf( outfile, "\txorl %%eax,%%eax\n" );
1016     fprintf( outfile, "\tmovw %%cs,%%ax\n" );
1017     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegCs) - STACK_SPACE );
1018     fprintf( outfile, "\tmovw %%es,%%ax\n" );
1019     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegEs) - STACK_SPACE );
1020     fprintf( outfile, "\tmovw %%fs,%%ax\n" );
1021     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegFs) - STACK_SPACE );
1022     fprintf( outfile, "\tmovw %%gs,%%ax\n" );
1023     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegGs) - STACK_SPACE );
1024     fprintf( outfile, "\tmovw %%ss,%%ax\n" );
1025     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegSs) - STACK_SPACE );
1026     fprintf( outfile, "\tmovw %%ds,%%ax\n" );
1027     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegDs) - STACK_SPACE );
1028     fprintf( outfile, "\tmovw %%ax,%%es\n" );  /* set %es equal to %ds just in case */
1029
1030     fprintf( outfile, "\tmovl $0x%x,%%eax\n", CONTEXT86_FULL );
1031     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(ContextFlags) - STACK_SPACE );
1032
1033     fprintf( outfile, "\tmovl 8(%%ebp),%%eax\n" ); /* Get %eip at time of call */
1034     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Eip) - STACK_SPACE );
1035
1036     /* Transfer the arguments */
1037
1038     fprintf( outfile, "\tmovl 4(%%ebp),%%ebx\n" );   /* get relay code addr */
1039     fprintf( outfile, "\tpushl %%esp\n" );           /* push ptr to context struct */
1040     fprintf( outfile, "\tmovzbl 4(%%ebx),%%ecx\n" ); /* fetch number of args to copy */
1041     fprintf( outfile, "\tjecxz 1f\n" );
1042     fprintf( outfile, "\tsubl %%ecx,%%esp\n" );
1043     fprintf( outfile, "\tleal 12(%%ebp),%%esi\n" );  /* get %esp at time of call */
1044     fprintf( outfile, "\tmovl %%esp,%%edi\n" );
1045     fprintf( outfile, "\tshrl $2,%%ecx\n" );
1046     fprintf( outfile, "\tcld\n" );
1047     fprintf( outfile, "\trep\n\tmovsl\n" );  /* copy args */
1048
1049     fprintf( outfile, "1:\tmovzbl 5(%%ebx),%%eax\n" ); /* fetch number of args to remove */
1050     fprintf( outfile, "\tleal 12(%%ebp,%%eax),%%eax\n" );
1051     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Esp) - STACK_SPACE );
1052
1053     /* Call the entry point */
1054
1055     fprintf( outfile, "\tcall *0(%%ebx)\n" );
1056
1057     /* Store %eip and %ebp onto the new stack */
1058
1059     fprintf( outfile, "\tmovl %d(%%ebp),%%edx\n", CONTEXTOFFSET(Esp) - STACK_SPACE );
1060     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(Eip) - STACK_SPACE );
1061     fprintf( outfile, "\tmovl %%eax,-4(%%edx)\n" );
1062     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(Ebp) - STACK_SPACE );
1063     fprintf( outfile, "\tmovl %%eax,-8(%%edx)\n" );
1064
1065     /* Restore the context structure */
1066
1067     /* Note: we don't bother to restore %cs, %ds and %ss
1068      *       changing them in 32-bit code is a recipe for disaster anyway
1069      */
1070     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(SegEs) - STACK_SPACE );
1071     fprintf( outfile, "\tmovw %%ax,%%es\n" );
1072     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(SegFs) - STACK_SPACE );
1073     fprintf( outfile, "\tmovw %%ax,%%fs\n" );
1074     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(SegGs) - STACK_SPACE );
1075     fprintf( outfile, "\tmovw %%ax,%%gs\n" );
1076
1077     fprintf( outfile, "\tmovl %d(%%ebp),%%edi\n", CONTEXTOFFSET(Edi) - STACK_SPACE );
1078     fprintf( outfile, "\tmovl %d(%%ebp),%%esi\n", CONTEXTOFFSET(Esi) - STACK_SPACE );
1079     fprintf( outfile, "\tmovl %d(%%ebp),%%edx\n", CONTEXTOFFSET(Edx) - STACK_SPACE );
1080     fprintf( outfile, "\tmovl %d(%%ebp),%%ecx\n", CONTEXTOFFSET(Ecx) - STACK_SPACE );
1081     fprintf( outfile, "\tmovl %d(%%ebp),%%ebx\n", CONTEXTOFFSET(Ebx) - STACK_SPACE );
1082
1083     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(EFlags) - STACK_SPACE );
1084     fprintf( outfile, "\tpushl %%eax\n" );
1085     fprintf( outfile, "\tpopfl\n" );
1086     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(Eax) - STACK_SPACE );
1087
1088     fprintf( outfile, "\tmovl %d(%%ebp),%%ebp\n", CONTEXTOFFSET(Esp) - STACK_SPACE );
1089     fprintf( outfile, "\tleal -8(%%ebp),%%esp\n" );
1090     fprintf( outfile, "\tpopl %%ebp\n" );
1091     fprintf( outfile, "\tret\n" );
1092 }
1093
1094
1095 /*******************************************************************
1096  *         BuildRelays16
1097  *
1098  * Build all the 16-bit relay callbacks
1099  */
1100 void BuildRelays16( FILE *outfile )
1101 {
1102     char buffer[1024];
1103
1104     /* File header */
1105
1106     fprintf( outfile, "/* File generated automatically. Do not edit! */\n\n" );
1107     fprintf( outfile, "\t.text\n" );
1108
1109     if (output_file_name && use_stabs)
1110     {
1111         getcwd(buffer, sizeof(buffer));
1112         fprintf( outfile, "\t.file\t\"%s\"\n", output_file_name );
1113
1114         /*
1115          * The stabs help the internal debugger as they are an indication that it
1116          * is sensible to step into a thunk/trampoline.
1117          */
1118         fprintf( outfile, ".stabs \"%s/\",100,0,0,Code_Start\n", buffer);
1119         fprintf( outfile, ".stabs \"%s\",100,0,0,Code_Start\n", output_file_name );
1120         fprintf( outfile, "Code_Start:\n\n" );
1121     }
1122
1123     fprintf( outfile, __ASM_NAME("Call16_Start") ":\n" );
1124     fprintf( outfile, "\t.globl " __ASM_NAME("Call16_Start") "\n" );
1125     fprintf( outfile, "\t.byte 0\n\n" );
1126
1127     /* Standard CallFrom16 routine (WORD return) */
1128     BuildCallFrom16Core( outfile, FALSE, FALSE, TRUE );
1129
1130     /* Standard CallFrom16 routine (DWORD return) */
1131     BuildCallFrom16Core( outfile, FALSE, FALSE, FALSE );
1132
1133     /* Register CallFrom16 routine */
1134     BuildCallFrom16Core( outfile, TRUE, FALSE, FALSE );
1135
1136     /* C16ThkSL CallFrom16 routine */
1137     BuildCallFrom16Core( outfile, FALSE, TRUE, FALSE );
1138
1139     /* Standard CallTo16 routine */
1140     BuildCallTo16Core( outfile, 0 );
1141
1142     /* Register CallTo16 routine (16:16 retf) */
1143     BuildCallTo16Core( outfile, 1 );
1144
1145     /* Register CallTo16 routine (16:32 retf) */
1146     BuildCallTo16Core( outfile, 2 );
1147
1148     /* CBClientThunkSL routine */
1149     BuildCallTo32CBClient( outfile, FALSE );
1150
1151     /* CBClientThunkSLEx routine */
1152     BuildCallTo32CBClient( outfile, TRUE  );
1153
1154     fprintf( outfile, __ASM_NAME("Call16_End") ":\n" );
1155     fprintf( outfile, "\t.globl " __ASM_NAME("Call16_End") "\n" );
1156
1157     if (use_stabs)
1158     {
1159         fprintf( outfile, "\t.stabs \"\",100,0,0,.Letext\n");
1160         fprintf( outfile, ".Letext:\n");
1161
1162         /* Some versions of gdb need to have the filename data for
1163            each section, so output it again for the data section. */
1164         if (output_file_name)
1165         {
1166             fprintf( outfile, ".stabs \"%s/\",100,0,0,Data_Start\n", buffer);
1167             fprintf( outfile, ".stabs \"%s\",100,0,0,Data_Start\n", output_file_name );
1168             fprintf( outfile, "Data_Start:\n\n" );
1169         }
1170     }
1171
1172     /* The whole Call16_Ret segment must lie within the .data section */
1173     fprintf( outfile, "\n\t.data\n" );
1174     fprintf( outfile, "\t.globl " __ASM_NAME("Call16_Ret_Start") "\n" );
1175     fprintf( outfile, __ASM_NAME("Call16_Ret_Start") ":\n" );
1176
1177     /* Standard CallTo16 return stub */
1178     BuildRet16Func( outfile );
1179
1180     /* CBClientThunkSL return stub */
1181     BuildCallTo32CBClientRet( outfile, FALSE );
1182
1183     /* CBClientThunkSLEx return stub */
1184     BuildCallTo32CBClientRet( outfile, TRUE  );
1185
1186     /* End of Call16_Ret segment */
1187     fprintf( outfile, "\n\t.globl " __ASM_NAME("Call16_Ret_End") "\n" );
1188     fprintf( outfile, __ASM_NAME("Call16_Ret_End") ":\n" );
1189 }
1190
1191 /*******************************************************************
1192  *         BuildRelays32
1193  *
1194  * Build all the 32-bit relay callbacks
1195  */
1196 void BuildRelays32( FILE *outfile )
1197 {
1198     /* File header */
1199
1200     fprintf( outfile, "/* File generated automatically. Do not edit! */\n\n" );
1201     fprintf( outfile, "\t.text\n" );
1202
1203     if (output_file_name && use_stabs)
1204     {
1205         char buffer[1024];
1206         getcwd(buffer, sizeof(buffer));
1207         fprintf( outfile, "\t.file\t\"%s\"\n", output_file_name );
1208
1209         /*
1210          * The stabs help the internal debugger as they are an indication that it
1211          * is sensible to step into a thunk/trampoline.
1212          */
1213         fprintf( outfile, ".stabs \"%s/\",100,0,0,Code_Start\n", buffer);
1214         fprintf( outfile, ".stabs \"%s\",100,0,0,Code_Start\n", output_file_name );
1215         fprintf( outfile, "Code_Start:\n\n" );
1216     }
1217
1218     /* 32-bit register entry point */
1219     BuildCallFrom32Regs( outfile );
1220
1221     if (use_stabs)
1222     {
1223         fprintf( outfile, "\t.stabs \"\",100,0,0,.Letext\n");
1224         fprintf( outfile, ".Letext:\n");
1225     }
1226 }
1227
1228 #else /* __i386__ */
1229
1230 void BuildRelays16( FILE *outfile )
1231 {
1232     fprintf( outfile, "/* File not used with this architecture. Do not edit! */\n\n" );
1233 }
1234
1235 void BuildRelays32( FILE *outfile )
1236 {
1237     fprintf( outfile, "/* File not used with this architecture. Do not edit! */\n\n" );
1238 }
1239
1240 #endif  /* __i386__ */