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 #ifdef __i386__
  36
  37 static void function_header( FILE *outfile, const char *name )
  38 {
  39     fprintf( outfile, "\n\t.align %d\n", get_alignment(4) );
  40     fprintf( outfile, "\t" __ASM_FUNC("%s") "\n", name );
  41     fprintf( outfile, "\t.globl " __ASM_NAME("%s") "\n", name );
  42     fprintf( outfile, __ASM_NAME("%s") ":\n", name );
  43 }
  44
  45
  46 static void function_footer( FILE *outfile, const char *name )
  47 {
  48     fprintf( outfile, ".size " __ASM_NAME("%s") ", . - " __ASM_NAME("%s") "\n", name, name );
  49 }
  50
  51 /*******************************************************************
  52  *         BuildCallFrom16Core
  53  *
  54  * This routine builds the core routines used in 16->32 thunks:
  55  * CallFrom16Word, CallFrom16Long, CallFrom16Register, and CallFrom16Thunk.
  56  *
  57  * These routines are intended to be called via a far call (with 32-bit
  58  * operand size) from 16-bit code.  The 16-bit code stub must push %bp,
  59  * the 32-bit entry point to be called, and the argument conversion
  60  * routine to be used (see stack layout below).
  61  *
  62  * The core routine completes the STACK16FRAME on the 16-bit stack and
  63  * switches to the 32-bit stack.  Then, the argument conversion routine
  64  * is called; it gets passed the 32-bit entry point and a pointer to the
  65  * 16-bit arguments (on the 16-bit stack) as parameters. (You can either
  66  * use conversion routines automatically generated by BuildCallFrom16,
  67  * or write your own for special purposes.)
  68  *
  69  * The conversion routine must call the 32-bit entry point, passing it
  70  * the converted arguments, and return its return value to the core.
  71  * After the conversion routine has returned, the core switches back
  72  * to the 16-bit stack, converts the return value to the DX:AX format
  73  * (CallFrom16Long), and returns to the 16-bit call stub.  All parameters,
  74  * including %bp, are popped off the stack.
  75  *
  76  * The 16-bit call stub now returns to the caller, popping the 16-bit
  77  * arguments if necessary (pascal calling convention).
  78  *
  79  * In the case of a 'register' function, CallFrom16Register fills a
  80  * CONTEXT86 structure with the values all registers had at the point
  81  * the first instruction of the 16-bit call stub was about to be
  82  * executed.  A pointer to this CONTEXT86 is passed as third parameter
  83  * to the argument conversion routine, which typically passes it on
  84  * to the called 32-bit entry point.
  85  *
  86  * CallFrom16Thunk is a special variant used by the implementation of
  87  * the Win95 16->32 thunk functions C16ThkSL and C16ThkSL01 and is
  88  * implemented as follows:
  89  * On entry, the EBX register is set up to contain a flat pointer to the
  90  * 16-bit stack such that EBX+22 points to the first argument.
  91  * Then, the entry point is called, while EBP is set up to point
  92  * to the return address (on the 32-bit stack).
  93  * The called function returns with CX set to the number of bytes
  94  * to be popped of the caller's stack.
  95  *
  96  * Stack layout upon entry to the core routine (STACK16FRAME):
  97  *  ...           ...
  98  * (sp+24) word   first 16-bit arg
  99  * (sp+22) word   cs
 100  * (sp+20) word   ip
 101  * (sp+18) word   bp
 102  * (sp+14) long   32-bit entry point (reused for Win16 mutex recursion count)
 103  * (sp+12) word   ip of actual entry point (necessary for relay debugging)
 104  * (sp+8)  long   relay (argument conversion) function entry point
 105  * (sp+4)  long   cs of 16-bit entry point
 106  * (sp)    long   ip of 16-bit entry point
 107  *
 108  * Added on the stack:
 109  * (sp-2)  word   saved gs
 110  * (sp-4)  word   saved fs
 111  * (sp-6)  word   saved es
 112  * (sp-8)  word   saved ds
 113  * (sp-12) long   saved ebp
 114  * (sp-16) long   saved ecx
 115  * (sp-20) long   saved edx
 116  * (sp-24) long   saved previous stack
 117  */
 118 static void BuildCallFrom16Core( FILE *outfile, int reg_func, int thunk, int short_ret )
 119 {
 120     const char *name = thunk? "thunk" : reg_func? "regs" : short_ret? "word" : "long";
 121
 122     /* Function header */
 123     if (thunk) function_header( outfile, "__wine_call_from_16_thunk" );
 124     else if (reg_func) function_header( outfile, "__wine_call_from_16_regs" );
 125     else if (short_ret) function_header( outfile, "__wine_call_from_16_word" );
 126     else function_header( outfile, "__wine_call_from_16_long" );
 127
 128     /* Create STACK16FRAME (except STACK32FRAME link) */
 129     fprintf( outfile, "\tpushw %%gs\n" );
 130     fprintf( outfile, "\tpushw %%fs\n" );
 131     fprintf( outfile, "\tpushw %%es\n" );
 132     fprintf( outfile, "\tpushw %%ds\n" );
 133     fprintf( outfile, "\tpushl %%ebp\n" );
 134     fprintf( outfile, "\tpushl %%ecx\n" );
 135     fprintf( outfile, "\tpushl %%edx\n" );
 136
 137     /* Save original EFlags register */
 138     fprintf( outfile, "\tpushfl\n" );
 139
 140     if ( UsePIC )
 141     {
 142         /* Get Global Offset Table into %ecx */
 143         fprintf( outfile, "\tcall .L__wine_call_from_16_%s.getgot1\n", name );
 144         fprintf( outfile, ".L__wine_call_from_16_%s.getgot1:\n", name );
 145         fprintf( outfile, "\tpopl %%ecx\n" );
 146         fprintf( outfile, "\taddl $_GLOBAL_OFFSET_TABLE_+[.-.L__wine_call_from_16_%s.getgot1], %%ecx\n", name );
 147     }
 148
 149     if (UsePIC)
 150     {
 151         fprintf( outfile, "\t.byte 0x2e\n\tmovl " __ASM_NAME("CallTo16_DataSelector@GOT") "(%%ecx), %%edx\n" );
 152         fprintf( outfile, "\t.byte 0x2e\n\tmovl (%%edx), %%edx\n" );
 153     }
 154     else
 155         fprintf( outfile, "\t.byte 0x2e\n\tmovl " __ASM_NAME("CallTo16_DataSelector") ",%%edx\n" );
 156
 157     /* Load 32-bit segment registers */
 158 #ifdef __svr4__
 159     fprintf( outfile, "\tdata16\n");
 160 #endif
 161     fprintf( outfile, "\tmovw %%dx, %%ds\n" );
 162 #ifdef __svr4__
 163     fprintf( outfile, "\tdata16\n");
 164 #endif
 165     fprintf( outfile, "\tmovw %%dx, %%es\n" );
 166
 167     if ( UsePIC )
 168     {
 169         fprintf( outfile, "\tmovl " __ASM_NAME("CallTo16_TebSelector@GOT") "(%%ecx), %%edx\n" );
 170         fprintf( outfile, "\tmovw (%%edx), %%fs\n" );
 171     }
 172     else
 173         fprintf( outfile, "\tmovw " __ASM_NAME("CallTo16_TebSelector") ", %%fs\n" );
 174
 175     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d),%%gs\n", STRUCTOFFSET(TEB,gs_sel) );
 176
 177     /* Get address of wine_ldt_copy array into %ecx */
 178     if ( UsePIC )
 179         fprintf( outfile, "\tmovl " __ASM_NAME("wine_ldt_copy@GOT") "(%%ecx), %%ecx\n" );
 180     else
 181         fprintf( outfile, "\tmovl $" __ASM_NAME("wine_ldt_copy") ", %%ecx\n" );
 182
 183     /* Translate STACK16FRAME base to flat offset in %edx */
 184     fprintf( outfile, "\tmovw %%ss, %%dx\n" );
 185     fprintf( outfile, "\tandl $0xfff8, %%edx\n" );
 186     fprintf( outfile, "\tshrl $1, %%edx\n" );
 187     fprintf( outfile, "\tmovl (%%ecx,%%edx), %%edx\n" );
 188     fprintf( outfile, "\tmovzwl %%sp, %%ebp\n" );
 189     fprintf( outfile, "\tleal (%%ebp,%%edx), %%edx\n" );
 190
 191     /* Get saved flags into %ecx */
 192     fprintf( outfile, "\tpopl %%ecx\n" );
 193
 194     /* Get the 32-bit stack pointer from the TEB and complete STACK16FRAME */
 195     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d), %%ebp\n", STACKOFFSET );
 196     fprintf( outfile, "\tpushl %%ebp\n" );
 197
 198     /* Switch stacks */
 199 #ifdef __svr4__
 200     fprintf( outfile,"\tdata16\n");
 201 #endif
 202     fprintf( outfile, "\t.byte 0x64\n\tmovw %%ss, (%d)\n", STACKOFFSET + 2 );
 203     fprintf( outfile, "\t.byte 0x64\n\tmovw %%sp, (%d)\n", STACKOFFSET );
 204     fprintf( outfile, "\tpushl %%ds\n" );
 205     fprintf( outfile, "\tpopl %%ss\n" );
 206     fprintf( outfile, "\tmovl %%ebp, %%esp\n" );
 207     fprintf( outfile, "\taddl $%d, %%ebp\n", STRUCTOFFSET(STACK32FRAME, ebp) );
 208
 209
 210     /* At this point:
 211        STACK16FRAME is completely set up
 212        DS, ES, SS: flat data segment
 213        FS: current TEB
 214        ESP: points to last STACK32FRAME
 215        EBP: points to ebp member of last STACK32FRAME
 216        EDX: points to current STACK16FRAME
 217        ECX: contains saved flags
 218        all other registers: unchanged */
 219
 220     /* Special case: C16ThkSL stub */
 221     if ( thunk )
 222     {
 223         /* Set up registers as expected and call thunk */
 224         fprintf( outfile, "\tleal %d(%%edx), %%ebx\n", sizeof(STACK16FRAME)-22 );
 225         fprintf( outfile, "\tleal -4(%%esp), %%ebp\n" );
 226
 227         fprintf( outfile, "\tcall *%d(%%edx)\n", STACK16OFFSET(entry_point) );
 228
 229         /* Switch stack back */
 230         fprintf( outfile, "\t.byte 0x64\n\tmovw (%d), %%ss\n", STACKOFFSET+2 );
 231         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%esp\n", STACKOFFSET );
 232         fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 233
 234         /* Restore registers and return directly to caller */
 235         fprintf( outfile, "\taddl $8, %%esp\n" );
 236         fprintf( outfile, "\tpopl %%ebp\n" );
 237         fprintf( outfile, "\tpopw %%ds\n" );
 238         fprintf( outfile, "\tpopw %%es\n" );
 239         fprintf( outfile, "\tpopw %%fs\n" );
 240         fprintf( outfile, "\tpopw %%gs\n" );
 241         fprintf( outfile, "\taddl $20, %%esp\n" );
 242
 243         fprintf( outfile, "\txorb %%ch, %%ch\n" );
 244         fprintf( outfile, "\tpopl %%ebx\n" );
 245         fprintf( outfile, "\taddw %%cx, %%sp\n" );
 246         fprintf( outfile, "\tpush %%ebx\n" );
 247
 248         fprintf( outfile, "\t.byte 0x66\n" );
 249         fprintf( outfile, "\tlret\n" );
 250
 251         return;
 252     }
 253
 254
 255     /* Build register CONTEXT */
 256     if ( reg_func )
 257     {
 258         fprintf( outfile, "\tsubl $%d, %%esp\n", sizeof(CONTEXT86) );
 259
 260         fprintf( outfile, "\tmovl %%ecx, %d(%%esp)\n", CONTEXTOFFSET(EFlags) );
 261
 262         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Eax) );
 263         fprintf( outfile, "\tmovl %%ebx, %d(%%esp)\n", CONTEXTOFFSET(Ebx) );
 264         fprintf( outfile, "\tmovl %%esi, %d(%%esp)\n", CONTEXTOFFSET(Esi) );
 265         fprintf( outfile, "\tmovl %%edi, %d(%%esp)\n", CONTEXTOFFSET(Edi) );
 266
 267         fprintf( outfile, "\tmovl %d(%%edx), %%eax\n", STACK16OFFSET(ebp) );
 268         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Ebp) );
 269         fprintf( outfile, "\tmovl %d(%%edx), %%eax\n", STACK16OFFSET(ecx) );
 270         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Ecx) );
 271         fprintf( outfile, "\tmovl %d(%%edx), %%eax\n", STACK16OFFSET(edx) );
 272         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Edx) );
 273
 274         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(ds) );
 275         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegDs) );
 276         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(es) );
 277         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegEs) );
 278         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(fs) );
 279         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegFs) );
 280         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(gs) );
 281         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegGs) );
 282
 283         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(cs) );
 284         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegCs) );
 285         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(ip) );
 286         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Eip) );
 287
 288         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%eax\n", STACKOFFSET+2 );
 289         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegSs) );
 290         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%eax\n", STACKOFFSET );
 291         fprintf( outfile, "\taddl $%d, %%eax\n", STACK16OFFSET(ip) );
 292         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Esp) );
 293 #if 0
 294         fprintf( outfile, "\tfsave %d(%%esp)\n", CONTEXTOFFSET(FloatSave) );
 295 #endif
 296
 297         /* Push address of CONTEXT86 structure -- popped by the relay routine */
 298         fprintf( outfile, "\tpushl %%esp\n" );
 299     }
 300
 301
 302     /* Print debug info before call */
 303     if ( debugging )
 304     {
 305         if ( UsePIC )
 306         {
 307             fprintf( outfile, "\tpushl %%ebx\n" );
 308
 309             /* Get Global Offset Table into %ebx (for PLT call) */
 310             fprintf( outfile, "\tcall .L__wine_call_from_16_%s.getgot2\n", name );
 311             fprintf( outfile, ".L__wine_call_from_16_%s.getgot2:\n", name );
 312             fprintf( outfile, "\tpopl %%ebx\n" );
 313             fprintf( outfile, "\taddl $_GLOBAL_OFFSET_TABLE_+[.-.L__wine_call_from_16_%s.getgot2], %%ebx\n", name );
 314         }
 315
 316         fprintf( outfile, "\tpushl %%edx\n" );
 317         if ( reg_func )
 318             fprintf( outfile, "\tleal -%d(%%ebp), %%eax\n\tpushl %%eax\n",
 319                               sizeof(CONTEXT) + STRUCTOFFSET(STACK32FRAME, ebp) );
 320         else
 321             fprintf( outfile, "\tpushl $0\n" );
 322
 323         if ( UsePIC )
 324             fprintf( outfile, "\tcall " __ASM_NAME("RELAY_DebugCallFrom16@PLT") "\n ");
 325         else
 326             fprintf( outfile, "\tcall " __ASM_NAME("RELAY_DebugCallFrom16") "\n ");
 327
 328         fprintf( outfile, "\tpopl %%edx\n" );
 329         fprintf( outfile, "\tpopl %%edx\n" );
 330
 331         if ( UsePIC )
 332             fprintf( outfile, "\tpopl %%ebx\n" );
 333     }
 334
 335     /* Call relay routine (which will call the API entry point) */
 336     fprintf( outfile, "\tleal %d(%%edx), %%eax\n", sizeof(STACK16FRAME) );
 337     fprintf( outfile, "\tpushl %%eax\n" );
 338     fprintf( outfile, "\tpushl %d(%%edx)\n", STACK16OFFSET(entry_point) );
 339     fprintf( outfile, "\tcall *%d(%%edx)\n", STACK16OFFSET(relay) );
 340
 341     /* Print debug info after call */
 342     if ( debugging )
 343     {
 344         if ( UsePIC )
 345         {
 346             fprintf( outfile, "\tpushl %%ebx\n" );
 347
 348             /* Get Global Offset Table into %ebx (for PLT call) */
 349             fprintf( outfile, "\tcall .L__wine_call_from_16_%s.getgot3\n", name );
 350             fprintf( outfile, ".L__wine_call_from_16_%s.getgot3:\n", name );
 351             fprintf( outfile, "\tpopl %%ebx\n" );
 352             fprintf( outfile, "\taddl $_GLOBAL_OFFSET_TABLE_+[.-.L__wine_call_from_16_%s.getgot3], %%ebx\n", name );
 353         }
 354
 355         fprintf( outfile, "\tpushl %%eax\n" );
 356         if ( reg_func )
 357             fprintf( outfile, "\tleal -%d(%%ebp), %%eax\n\tpushl %%eax\n",
 358                               sizeof(CONTEXT) + STRUCTOFFSET(STACK32FRAME, ebp) );
 359         else
 360             fprintf( outfile, "\tpushl $0\n" );
 361
 362         if ( UsePIC )
 363             fprintf( outfile, "\tcall " __ASM_NAME("RELAY_DebugCallFrom16Ret@PLT") "\n ");
 364         else
 365             fprintf( outfile, "\tcall " __ASM_NAME("RELAY_DebugCallFrom16Ret") "\n ");
 366
 367         fprintf( outfile, "\tpopl %%eax\n" );
 368         fprintf( outfile, "\tpopl %%eax\n" );
 369
 370         if ( UsePIC )
 371             fprintf( outfile, "\tpopl %%ebx\n" );
 372     }
 373
 374
 375     if ( reg_func )
 376     {
 377         fprintf( outfile, "\tmovl %%esp, %%ebx\n" );
 378
 379         /* Switch stack back */
 380         fprintf( outfile, "\t.byte 0x64\n\tmovw (%d), %%ss\n", STACKOFFSET+2 );
 381         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%esp\n", STACKOFFSET );
 382         fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 383
 384         /* Get return address to CallFrom16 stub */
 385         fprintf( outfile, "\taddw $%d, %%sp\n", STACK16OFFSET(callfrom_ip)-4 );
 386         fprintf( outfile, "\tpopl %%eax\n" );
 387         fprintf( outfile, "\tpopl %%edx\n" );
 388
 389         /* Restore all registers from CONTEXT */
 390         fprintf( outfile, "\tmovw %d(%%ebx), %%ss\n", CONTEXTOFFSET(SegSs) );
 391         fprintf( outfile, "\tmovl %d(%%ebx), %%esp\n", CONTEXTOFFSET(Esp) );
 392         fprintf( outfile, "\taddl $4, %%esp\n" );  /* room for final return address */
 393
 394         fprintf( outfile, "\tpushw %d(%%ebx)\n", CONTEXTOFFSET(SegCs) );
 395         fprintf( outfile, "\tpushw %d(%%ebx)\n", CONTEXTOFFSET(Eip) );
 396         fprintf( outfile, "\tpushl %%edx\n" );
 397         fprintf( outfile, "\tpushl %%eax\n" );
 398         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(EFlags) );
 399         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(SegDs) );
 400
 401         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(SegEs) );
 402         fprintf( outfile, "\tpopl %%es\n" );
 403         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(SegFs) );
 404         fprintf( outfile, "\tpopl %%fs\n" );
 405         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(SegGs) );
 406         fprintf( outfile, "\tpopl %%gs\n" );
 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     if (thunk) function_footer( outfile, "__wine_call_from_16_thunk" );
 445     else if (reg_func) function_footer( outfile, "__wine_call_from_16_regs" );
 446     else if (short_ret) function_footer( outfile, "__wine_call_from_16_word" );
 447     else function_footer( outfile, "__wine_call_from_16_long" );
 448 }
 449
 450
 451 /*******************************************************************
 452  *         BuildCallTo16Core
 453  *
 454  * This routine builds the core routines used in 32->16 thunks:
 455  *
 456  * extern DWORD WINAPI wine_call_to_16( FARPROC16 target, DWORD cbArgs, PEXCEPTION_HANDLER handler );
 457  * extern void WINAPI wine_call_to_16_regs( CONTEXT86 *context, DWORD cbArgs, PEXCEPTION_HANDLER handler );
 458  *
 459  * These routines can be called directly from 32-bit code.
 460  *
 461  * All routines expect that the 16-bit stack contents (arguments) and the
 462  * return address (segptr to CallTo16_Ret) were already set up by the
 463  * caller; nb_args must contain the number of bytes to be conserved.  The
 464  * 16-bit SS:SP will be set accordinly.
 465  *
 466  * All other registers are either taken from the CONTEXT86 structure
 467  * or else set to default values.  The target routine address is either
 468  * given directly or taken from the CONTEXT86.
 469  */
 470 static void BuildCallTo16Core( FILE *outfile, int reg_func )
 471 {
 472     const char *name = reg_func ? "wine_call_to_16_regs" : "wine_call_to_16";
 473
 474     /* Function header */
 475     function_header( outfile, name );
 476
 477     /* Function entry sequence */
 478     fprintf( outfile, "\tpushl %%ebp\n" );
 479     fprintf( outfile, "\tmovl %%esp, %%ebp\n" );
 480
 481     /* Save the 32-bit registers */
 482     fprintf( outfile, "\tpushl %%ebx\n" );
 483     fprintf( outfile, "\tpushl %%esi\n" );
 484     fprintf( outfile, "\tpushl %%edi\n" );
 485     fprintf( outfile, "\t.byte 0x64\n\tmovl %%gs,(%d)\n", STRUCTOFFSET(TEB,gs_sel) );
 486
 487     /* Setup exception frame */
 488     fprintf( outfile, "\t.byte 0x64\n\tpushl (%d)\n", STACKOFFSET );
 489     fprintf( outfile, "\tpushl 16(%%ebp)\n" ); /* handler */
 490     fprintf( outfile, "\t.byte 0x64\n\tpushl (%d)\n", STRUCTOFFSET(TEB,Tib.ExceptionList) );
 491     fprintf( outfile, "\t.byte 0x64\n\tmovl %%esp,(%d)\n", STRUCTOFFSET(TEB,Tib.ExceptionList) );
 492
 493     /* Call the actual CallTo16 routine (simulate a lcall) */
 494     fprintf( outfile, "\tpushl %%cs\n" );
 495     fprintf( outfile, "\tcall .L%s\n", name );
 496
 497     /* Remove exception frame */
 498     fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STRUCTOFFSET(TEB,Tib.ExceptionList) );
 499     fprintf( outfile, "\taddl $4, %%esp\n" );
 500     fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 501
 502     if ( !reg_func )
 503     {
 504         /* Convert return value */
 505         fprintf( outfile, "\tandl $0xffff,%%eax\n" );
 506         fprintf( outfile, "\tshll $16,%%edx\n" );
 507         fprintf( outfile, "\torl %%edx,%%eax\n" );
 508     }
 509     else
 510     {
 511         /*
 512          * Modify CONTEXT86 structure to contain new values
 513          *
 514          * NOTE:  We restore only EAX, EBX, EDX, EDX, EBP, and ESP.
 515          *        The segment registers as well as ESI and EDI should
 516          *        not be modified by a well-behaved 16-bit routine in
 517          *        any case.  [If necessary, we could restore them as well,
 518          *        at the cost of a somewhat less efficient return path.]
 519          */
 520
 521         fprintf( outfile, "\tmovl %d(%%esp), %%edi\n", STACK32OFFSET(target) - STACK32OFFSET(edi));
 522                 /* everything above edi has been popped already */
 523
 524         fprintf( outfile, "\tmovl %%eax, %d(%%edi)\n", CONTEXTOFFSET(Eax) );
 525         fprintf( outfile, "\tmovl %%ebx, %d(%%edi)\n", CONTEXTOFFSET(Ebx) );
 526         fprintf( outfile, "\tmovl %%ecx, %d(%%edi)\n", CONTEXTOFFSET(Ecx) );
 527         fprintf( outfile, "\tmovl %%edx, %d(%%edi)\n", CONTEXTOFFSET(Edx) );
 528         fprintf( outfile, "\tmovl %%ebp, %d(%%edi)\n", CONTEXTOFFSET(Ebp) );
 529         fprintf( outfile, "\tmovl %%esi, %d(%%edi)\n", CONTEXTOFFSET(Esp) );
 530                  /* The return glue code saved %esp into %esi */
 531     }
 532
 533     /* Restore the 32-bit registers */
 534     fprintf( outfile, "\tpopl %%edi\n" );
 535     fprintf( outfile, "\tpopl %%esi\n" );
 536     fprintf( outfile, "\tpopl %%ebx\n" );
 537
 538     /* Function exit sequence */
 539     fprintf( outfile, "\tpopl %%ebp\n" );
 540     fprintf( outfile, "\tret $12\n" );
 541
 542
 543     /* Start of the actual CallTo16 routine */
 544
 545     fprintf( outfile, ".L%s:\n", name );
 546
 547     /* Switch to the 16-bit stack */
 548     fprintf( outfile, "\tmovl %%esp,%%edx\n" );
 549 #ifdef __svr4__
 550     fprintf( outfile,"\tdata16\n");
 551 #endif
 552     fprintf( outfile, "\t.byte 0x64\n\tmovw (%d),%%ss\n", STACKOFFSET + 2);
 553     fprintf( outfile, "\t.byte 0x64\n\tmovw (%d),%%sp\n", STACKOFFSET );
 554     fprintf( outfile, "\t.byte 0x64\n\tmovl %%edx,(%d)\n", STACKOFFSET );
 555
 556     /* Make %bp point to the previous stackframe (built by CallFrom16) */
 557     fprintf( outfile, "\tmovzwl %%sp,%%ebp\n" );
 558     fprintf( outfile, "\tleal %d(%%ebp),%%ebp\n", STACK16OFFSET(bp) );
 559
 560     /* Add the specified offset to the new sp */
 561     fprintf( outfile, "\tsubw %d(%%edx), %%sp\n", STACK32OFFSET(nb_args) );
 562
 563     if (reg_func)
 564     {
 565         /* Push the called routine address */
 566         fprintf( outfile, "\tmovl %d(%%edx),%%edx\n", STACK32OFFSET(target) );
 567         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(SegCs) );
 568         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(Eip) );
 569
 570         /* Get the registers */
 571         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(SegDs) );
 572         fprintf( outfile, "\tpushl %d(%%edx)\n", CONTEXTOFFSET(SegEs) );
 573         fprintf( outfile, "\tpopl %%es\n" );
 574         fprintf( outfile, "\tpushl %d(%%edx)\n", CONTEXTOFFSET(SegFs) );
 575         fprintf( outfile, "\tpopl %%fs\n" );
 576         fprintf( outfile, "\tpushl %d(%%edx)\n", CONTEXTOFFSET(SegGs) );
 577         fprintf( outfile, "\tpopl %%gs\n" );
 578         fprintf( outfile, "\tmovl %d(%%edx),%%ebp\n", CONTEXTOFFSET(Ebp) );
 579         fprintf( outfile, "\tmovl %d(%%edx),%%esi\n", CONTEXTOFFSET(Esi) );
 580         fprintf( outfile, "\tmovl %d(%%edx),%%edi\n", CONTEXTOFFSET(Edi) );
 581         fprintf( outfile, "\tmovl %d(%%edx),%%eax\n", CONTEXTOFFSET(Eax) );
 582         fprintf( outfile, "\tmovl %d(%%edx),%%ebx\n", CONTEXTOFFSET(Ebx) );
 583         fprintf( outfile, "\tmovl %d(%%edx),%%ecx\n", CONTEXTOFFSET(Ecx) );
 584         fprintf( outfile, "\tmovl %d(%%edx),%%edx\n", CONTEXTOFFSET(Edx) );
 585
 586         /* Get the 16-bit ds */
 587         fprintf( outfile, "\tpopw %%ds\n" );
 588     }
 589     else  /* not a register function */
 590     {
 591         /* Push the called routine address */
 592         fprintf( outfile, "\tpushl %d(%%edx)\n", STACK32OFFSET(target) );
 593
 594         /* Set %fs and %gs to the value saved by the last CallFrom16 */
 595         fprintf( outfile, "\tpushw %d(%%ebp)\n", STACK16OFFSET(fs)-STACK16OFFSET(bp) );
 596         fprintf( outfile, "\tpopw %%fs\n" );
 597         fprintf( outfile, "\tpushw %d(%%ebp)\n", STACK16OFFSET(gs)-STACK16OFFSET(bp) );
 598         fprintf( outfile, "\tpopw %%gs\n" );
 599
 600         /* Set %ds and %es (and %ax just in case) equal to %ss */
 601         fprintf( outfile, "\tmovw %%ss,%%ax\n" );
 602         fprintf( outfile, "\tmovw %%ax,%%ds\n" );
 603         fprintf( outfile, "\tmovw %%ax,%%es\n" );
 604     }
 605
 606     /* Jump to the called routine */
 607     fprintf( outfile, "\t.byte 0x66\n" );
 608     fprintf( outfile, "\tlret\n" );
 609
 610     /* Function footer */
 611     function_footer( outfile, name );
 612 }
 613
 614
 615 /*******************************************************************
 616  *         BuildRet16Func
 617  *
 618  * Build the return code for 16-bit callbacks
 619  */
 620 static void BuildRet16Func( FILE *outfile )
 621 {
 622     function_header( outfile, "CallTo16_Ret" );
 623
 624     /* Save %esp into %esi */
 625     fprintf( outfile, "\tmovl %%esp,%%esi\n" );
 626
 627     /* Restore 32-bit segment registers */
 628
 629     fprintf( outfile, "\t.byte 0x2e\n\tmovl " __ASM_NAME("CallTo16_DataSelector") "-" __ASM_NAME("Call16_Ret_Start") ",%%edi\n" );
 630 #ifdef __svr4__
 631     fprintf( outfile, "\tdata16\n");
 632 #endif
 633     fprintf( outfile, "\tmovw %%di,%%ds\n" );
 634 #ifdef __svr4__
 635     fprintf( outfile, "\tdata16\n");
 636 #endif
 637     fprintf( outfile, "\tmovw %%di,%%es\n" );
 638
 639     fprintf( outfile, "\t.byte 0x2e\n\tmovl " __ASM_NAME("CallTo16_TebSelector") "-" __ASM_NAME("Call16_Ret_Start") ",%%fs\n" );
 640
 641     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d),%%gs\n", STRUCTOFFSET(TEB,gs_sel) );
 642
 643     /* Restore the 32-bit stack */
 644
 645 #ifdef __svr4__
 646     fprintf( outfile, "\tdata16\n");
 647 #endif
 648     fprintf( outfile, "\tmovw %%di,%%ss\n" );
 649     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d),%%esp\n", STACKOFFSET );
 650
 651     /* Return to caller */
 652
 653     fprintf( outfile, "\tlret\n" );
 654
 655     /* Function footer */
 656     function_footer( outfile, "CallTo16_Ret" );
 657
 658     /* Declare the return address and data selector variables */
 659
 660     fprintf( outfile, "\n\t.align %d\n", get_alignment(4) );
 661     fprintf( outfile, "\t.globl " __ASM_NAME("CallTo16_DataSelector") "\n" );
 662     fprintf( outfile, __ASM_NAME("CallTo16_DataSelector") ":\t.long 0\n" );
 663     fprintf( outfile, "\t.globl " __ASM_NAME("CallTo16_TebSelector") "\n" );
 664     fprintf( outfile, __ASM_NAME("CallTo16_TebSelector") ":\t.long 0\n" );
 665 }
 666
 667
 668 /*******************************************************************
 669  *         BuildCallTo32CBClient
 670  *
 671  * Call a CBClient relay stub from 32-bit code (KERNEL.620).
 672  *
 673  * Since the relay stub is itself 32-bit, this should not be a problem;
 674  * unfortunately, the relay stubs are expected to switch back to a
 675  * 16-bit stack (and 16-bit code) after completion :-(
 676  *
 677  * This would conflict with our 16- vs. 32-bit stack handling, so
 678  * we simply switch *back* to our 32-bit stack before returning to
 679  * the caller ...
 680  *
 681  * The CBClient relay stub expects to be called with the following
 682  * 16-bit stack layout, and with ebp and ebx pointing into the 16-bit
 683  * stack at the designated places:
 684  *
 685  *    ...
 686  *  (ebp+14) original arguments to the callback routine
 687  *  (ebp+10) far return address to original caller
 688  *  (ebp+6)  Thunklet target address
 689  *  (ebp+2)  Thunklet relay ID code
 690  *  (ebp)    BP (saved by CBClientGlueSL)
 691  *  (ebp-2)  SI (saved by CBClientGlueSL)
 692  *  (ebp-4)  DI (saved by CBClientGlueSL)
 693  *  (ebp-6)  DS (saved by CBClientGlueSL)
 694  *
 695  *   ...     buffer space used by the 16-bit side glue for temp copies
 696  *
 697  *  (ebx+4)  far return address to 16-bit side glue code
 698  *  (ebx)    saved 16-bit ss:sp (pointing to ebx+4)
 699  *
 700  * The 32-bit side glue code accesses both the original arguments (via ebp)
 701  * and the temporary copies prepared by the 16-bit side glue (via ebx).
 702  * After completion, the stub will load ss:sp from the buffer at ebx
 703  * and perform a far return to 16-bit code.
 704  *
 705  * To trick the relay stub into returning to us, we replace the 16-bit
 706  * return address to the glue code by a cs:ip pair pointing to our
 707  * return entry point (the original return address is saved first).
 708  * Our return stub thus called will then reload the 32-bit ss:esp and
 709  * return to 32-bit code (by using and ss:esp value that we have also
 710  * pushed onto the 16-bit stack before and a cs:eip values found at
 711  * that position on the 32-bit stack).  The ss:esp to be restored is
 712  * found relative to the 16-bit stack pointer at:
 713  *
 714  *  (ebx-4)   ss  (flat)
 715  *  (ebx-8)   sp  (32-bit stack pointer)
 716  *
 717  * The second variant of this routine, CALL32_CBClientEx, which is used
 718  * to implement KERNEL.621, has to cope with yet another problem: Here,
 719  * the 32-bit side directly returns to the caller of the CBClient thunklet,
 720  * restoring registers saved by CBClientGlueSL and cleaning up the stack.
 721  * As we have to return to our 32-bit code first, we have to adapt the
 722  * layout of our temporary area so as to include values for the registers
 723  * that are to be restored, and later (in the implementation of KERNEL.621)
 724  * we *really* restore them. The return stub restores DS, DI, SI, and BP
 725  * from the stack, skips the next 8 bytes (CBClient relay code / target),
 726  * and then performs a lret NN, where NN is the number of arguments to be
 727  * removed. Thus, we prepare our temporary area as follows:
 728  *
 729  *     (ebx+22) 16-bit cs  (this segment)
 730  *     (ebx+20) 16-bit ip  ('16-bit' return entry point)
 731  *     (ebx+16) 32-bit ss  (flat)
 732  *     (ebx+12) 32-bit sp  (32-bit stack pointer)
 733  *     (ebx+10) 16-bit bp  (points to ebx+24)
 734  *     (ebx+8)  16-bit si  (ignored)
 735  *     (ebx+6)  16-bit di  (ignored)
 736  *     (ebx+4)  16-bit ds  (we actually use the flat DS here)
 737  *     (ebx+2)  16-bit ss  (16-bit stack segment)
 738  *     (ebx+0)  16-bit sp  (points to ebx+4)
 739  *
 740  * Note that we ensure that DS is not changed and remains the flat segment,
 741  * and the 32-bit stack pointer our own return stub needs fits just
 742  * perfectly into the 8 bytes that are skipped by the Windows stub.
 743  * One problem is that we have to determine the number of removed arguments,
 744  * as these have to be really removed in KERNEL.621. Thus, the BP value
 745  * that we place in the temporary area to be restored, contains the value
 746  * that SP would have if no arguments were removed. By comparing the actual
 747  * value of SP with this value in our return stub we can compute the number
 748  * of removed arguments. This is then returned to KERNEL.621.
 749  *
 750  * The stack layout of this function:
 751  * (ebp+20)  nArgs     pointer to variable receiving nr. of args (Ex only)
 752  * (ebp+16)  esi       pointer to caller's esi value
 753  * (ebp+12)  arg       ebp value to be set for relay stub
 754  * (ebp+8)   func      CBClient relay stub address
 755  * (ebp+4)   ret addr
 756  * (ebp)     ebp
 757  */
 758 static void BuildCallTo32CBClient( FILE *outfile, BOOL isEx )
 759 {
 760     const char *name = isEx? "CBClientEx" : "CBClient";
 761     int size = isEx? 24 : 12;
 762
 763     /* Function header */
 764
 765     fprintf( outfile, "\n\t.align %d\n", get_alignment(4) );
 766     fprintf( outfile, "\t.globl " __ASM_NAME("CALL32_%s") "\n", name );
 767     fprintf( outfile, __ASM_NAME("CALL32_%s") ":\n", name );
 768
 769     /* Entry code */
 770
 771     fprintf( outfile, "\tpushl %%ebp\n" );
 772     fprintf( outfile, "\tmovl %%esp,%%ebp\n" );
 773     fprintf( outfile, "\tpushl %%edi\n" );
 774     fprintf( outfile, "\tpushl %%esi\n" );
 775     fprintf( outfile, "\tpushl %%ebx\n" );
 776
 777     /* Get the 16-bit stack */
 778
 779     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d),%%ebx\n", STACKOFFSET);
 780
 781     /* Convert it to a flat address */
 782
 783     fprintf( outfile, "\tshldl $16,%%ebx,%%eax\n" );
 784     fprintf( outfile, "\tandl $0xfff8,%%eax\n" );
 785     fprintf( outfile, "\tshrl $1,%%eax\n" );
 786     fprintf( outfile, "\tmovl " __ASM_NAME("wine_ldt_copy") "(%%eax),%%esi\n" );
 787     fprintf( outfile, "\tmovw %%bx,%%ax\n" );
 788     fprintf( outfile, "\taddl %%eax,%%esi\n" );
 789
 790     /* Allocate temporary area (simulate STACK16_PUSH) */
 791
 792     fprintf( outfile, "\tpushf\n" );
 793     fprintf( outfile, "\tcld\n" );
 794     fprintf( outfile, "\tleal -%d(%%esi), %%edi\n", size );
 795     fprintf( outfile, "\tmovl $%d, %%ecx\n", sizeof(STACK16FRAME) );
 796     fprintf( outfile, "\trep\n\tmovsb\n" );
 797     fprintf( outfile, "\tpopf\n" );
 798
 799     fprintf( outfile, "\t.byte 0x64\n\tsubw $%d,(%d)\n", size, STACKOFFSET );
 800
 801     fprintf( outfile, "\tpushl %%edi\n" );  /* remember address */
 802
 803     /* Set up temporary area */
 804
 805     if ( !isEx )
 806     {
 807         fprintf( outfile, "\tleal 4(%%edi), %%edi\n" );
 808
 809         fprintf( outfile, "\tleal -8(%%esp), %%eax\n" );
 810         fprintf( outfile, "\tmovl %%eax, -8(%%edi)\n" );    /* 32-bit sp */
 811
 812         fprintf( outfile, "\tmovw %%ss, %%ax\n" );
 813         fprintf( outfile, "\tandl $0x0000ffff, %%eax\n" );
 814         fprintf( outfile, "\tmovl %%eax, -4(%%edi)\n" );    /* 32-bit ss */
 815
 816         fprintf( outfile, "\taddl $%d, %%ebx\n", sizeof(STACK16FRAME)-size+4 + 4 );
 817         fprintf( outfile, "\tmovl %%ebx, 0(%%edi)\n" );    /* 16-bit ss:sp */
 818
 819         fprintf( outfile, "\tmovl " __ASM_NAME("CALL32_%s_RetAddr") ", %%eax\n", name );
 820         fprintf( outfile, "\tmovl %%eax, 4(%%edi)\n" );   /* overwrite return address */
 821     }
 822     else
 823     {
 824         fprintf( outfile, "\taddl $%d, %%ebx\n", sizeof(STACK16FRAME)-size+4 );
 825         fprintf( outfile, "\tmovl %%ebx, 0(%%edi)\n" );
 826
 827         fprintf( outfile, "\tmovw %%ds, %%ax\n" );
 828         fprintf( outfile, "\tmovw %%ax, 4(%%edi)\n" );
 829
 830         fprintf( outfile, "\taddl $20, %%ebx\n" );
 831         fprintf( outfile, "\tmovw %%bx, 10(%%edi)\n" );
 832
 833         fprintf( outfile, "\tleal -8(%%esp), %%eax\n" );
 834         fprintf( outfile, "\tmovl %%eax, 12(%%edi)\n" );
 835
 836         fprintf( outfile, "\tmovw %%ss, %%ax\n" );
 837         fprintf( outfile, "\tandl $0x0000ffff, %%eax\n" );
 838         fprintf( outfile, "\tmovl %%eax, 16(%%edi)\n" );
 839
 840         fprintf( outfile, "\tmovl " __ASM_NAME("CALL32_%s_RetAddr") ", %%eax\n", name );
 841         fprintf( outfile, "\tmovl %%eax, 20(%%edi)\n" );
 842     }
 843
 844     /* Set up registers and call CBClient relay stub (simulating a far call) */
 845
 846     fprintf( outfile, "\tmovl 16(%%ebp), %%esi\n" );
 847     fprintf( outfile, "\tmovl (%%esi), %%esi\n" );
 848
 849     fprintf( outfile, "\tmovl %%edi, %%ebx\n" );
 850     fprintf( outfile, "\tmovl 8(%%ebp), %%eax\n" );
 851     fprintf( outfile, "\tmovl 12(%%ebp), %%ebp\n" );
 852
 853     fprintf( outfile, "\tpushl %%cs\n" );
 854     fprintf( outfile, "\tcall *%%eax\n" );
 855
 856     /* Return new esi value to caller */
 857
 858     fprintf( outfile, "\tmovl 32(%%esp), %%edi\n" );
 859     fprintf( outfile, "\tmovl %%esi, (%%edi)\n" );
 860
 861     /* Cleanup temporary area (simulate STACK16_POP) */
 862
 863     fprintf( outfile, "\tpop %%esi\n" );
 864
 865     fprintf( outfile, "\tpushf\n" );
 866     fprintf( outfile, "\tstd\n" );
 867     fprintf( outfile, "\tdec %%esi\n" );
 868     fprintf( outfile, "\tleal %d(%%esi), %%edi\n", size );
 869     fprintf( outfile, "\tmovl $%d, %%ecx\n", sizeof(STACK16FRAME) );
 870     fprintf( outfile, "\trep\n\tmovsb\n" );
 871     fprintf( outfile, "\tpopf\n" );
 872
 873     fprintf( outfile, "\t.byte 0x64\n\taddw $%d,(%d)\n", size, STACKOFFSET );
 874
 875     /* Return argument size to caller */
 876     if ( isEx )
 877     {
 878         fprintf( outfile, "\tmovl 32(%%esp), %%ebx\n" );
 879         fprintf( outfile, "\tmovl %%ebp, (%%ebx)\n" );
 880     }
 881
 882     /* Restore registers and return */
 883
 884     fprintf( outfile, "\tpopl %%ebx\n" );
 885     fprintf( outfile, "\tpopl %%esi\n" );
 886     fprintf( outfile, "\tpopl %%edi\n" );
 887     fprintf( outfile, "\tpopl %%ebp\n" );
 888     fprintf( outfile, "\tret\n" );
 889     fprintf( outfile, ".size " __ASM_NAME("CALL32_%s") ", . - " __ASM_NAME("CALL32_%s") "\n", name, name );
 890 }
 891
 892 static void BuildCallTo32CBClientRet( FILE *outfile, BOOL isEx )
 893 {
 894     const char *name = isEx? "CBClientEx" : "CBClient";
 895
 896     /* '16-bit' return stub */
 897
 898     fprintf( outfile, "\n\t.globl " __ASM_NAME("CALL32_%s_Ret") "\n", name );
 899     fprintf( outfile, __ASM_NAME("CALL32_%s_Ret") ":\n", name );
 900
 901     if ( !isEx )
 902     {
 903         fprintf( outfile, "\tmovzwl %%sp, %%ebx\n" );
 904         fprintf( outfile, "\tlssl %%ss:-16(%%ebx), %%esp\n" );
 905     }
 906     else
 907     {
 908         fprintf( outfile, "\tmovzwl %%bp, %%ebx\n" );
 909         fprintf( outfile, "\tsubw %%bp, %%sp\n" );
 910         fprintf( outfile, "\tmovzwl %%sp, %%ebp\n" );
 911         fprintf( outfile, "\tlssl %%ss:-12(%%ebx), %%esp\n" );
 912     }
 913     fprintf( outfile, "\tlret\n" );
 914
 915     fprintf( outfile, ".size " __ASM_NAME("CALL32_%s_Ret") ", . - " __ASM_NAME("CALL32_%s_Ret") "\n", name, name );
 916
 917     /* Declare the return address variable */
 918
 919     fprintf( outfile, "\n\t.globl " __ASM_NAME("CALL32_%s_RetAddr") "\n", name );
 920     fprintf( outfile, __ASM_NAME("CALL32_%s_RetAddr") ":\t.long 0\n", name );
 921 }
 922
 923
 924 /*******************************************************************
 925  *         BuildCallFrom32Regs
 926  *
 927  * Build a 32-bit-to-Wine call-back function for a 'register' function.
 928  * 'args' is the number of dword arguments.
 929  *
 930  * Stack layout:
 931  *   ...
 932  * (ebp+12)  first arg
 933  * (ebp+8)   ret addr to user code
 934  * (ebp+4)   ret addr to relay code
 935  * (ebp+0)   saved ebp
 936  * (ebp-128) buffer area to allow stack frame manipulation
 937  * (ebp-332) CONTEXT86 struct
 938  * (ebp-336) CONTEXT86 *argument
 939  *  ....     other arguments copied from (ebp+12)
 940  *
 941  * The entry point routine is called with a CONTEXT* extra argument,
 942  * following the normal args. In this context structure, EIP_reg
 943  * contains the return address to user code, and ESP_reg the stack
 944  * pointer on return (with the return address and arguments already
 945  * removed).
 946  */
 947 static void BuildCallFrom32Regs( FILE *outfile )
 948 {
 949     static const int STACK_SPACE = 128 + sizeof(CONTEXT86);
 950
 951     /* Function header */
 952
 953     function_header( outfile, "__wine_call_from_32_regs" );
 954
 955     /* Allocate some buffer space on the stack */
 956
 957     fprintf( outfile, "\tpushl %%ebp\n" );
 958     fprintf( outfile, "\tmovl %%esp,%%ebp\n ");
 959     fprintf( outfile, "\tleal -%d(%%esp), %%esp\n", STACK_SPACE );
 960
 961     /* Build the context structure */
 962
 963     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Eax) - STACK_SPACE );
 964     fprintf( outfile, "\tpushfl\n" );
 965     fprintf( outfile, "\tpopl %%eax\n" );
 966     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(EFlags) - STACK_SPACE );
 967     fprintf( outfile, "\tmovl 0(%%ebp),%%eax\n" );
 968     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Ebp) - STACK_SPACE );
 969     fprintf( outfile, "\tmovl %%ebx,%d(%%ebp)\n", CONTEXTOFFSET(Ebx) - STACK_SPACE );
 970     fprintf( outfile, "\tmovl %%ecx,%d(%%ebp)\n", CONTEXTOFFSET(Ecx) - STACK_SPACE );
 971     fprintf( outfile, "\tmovl %%edx,%d(%%ebp)\n", CONTEXTOFFSET(Edx) - STACK_SPACE );
 972     fprintf( outfile, "\tmovl %%esi,%d(%%ebp)\n", CONTEXTOFFSET(Esi) - STACK_SPACE );
 973     fprintf( outfile, "\tmovl %%edi,%d(%%ebp)\n", CONTEXTOFFSET(Edi) - STACK_SPACE );
 974
 975     fprintf( outfile, "\txorl %%eax,%%eax\n" );
 976     fprintf( outfile, "\tmovw %%cs,%%ax\n" );
 977     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegCs) - STACK_SPACE );
 978     fprintf( outfile, "\tmovw %%es,%%ax\n" );
 979     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegEs) - STACK_SPACE );
 980     fprintf( outfile, "\tmovw %%fs,%%ax\n" );
 981     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegFs) - STACK_SPACE );
 982     fprintf( outfile, "\tmovw %%gs,%%ax\n" );
 983     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegGs) - STACK_SPACE );
 984     fprintf( outfile, "\tmovw %%ss,%%ax\n" );
 985     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegSs) - STACK_SPACE );
 986     fprintf( outfile, "\tmovw %%ds,%%ax\n" );
 987     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegDs) - STACK_SPACE );
 988     fprintf( outfile, "\tmovw %%ax,%%es\n" );  /* set %es equal to %ds just in case */
 989
 990     fprintf( outfile, "\tmovl $0x%x,%%eax\n", CONTEXT86_FULL );
 991     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(ContextFlags) - STACK_SPACE );
 992
 993     fprintf( outfile, "\tmovl 8(%%ebp),%%eax\n" ); /* Get %eip at time of call */
 994     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Eip) - STACK_SPACE );
 995
 996     /* Transfer the arguments */
 997
 998     fprintf( outfile, "\tmovl 4(%%ebp),%%ebx\n" );   /* get relay code addr */
 999     fprintf( outfile, "\tpushl %%esp\n" );           /* push ptr to context struct */
1000     fprintf( outfile, "\tmovzbl 4(%%ebx),%%ecx\n" ); /* fetch number of args to copy */
1001     fprintf( outfile, "\tjecxz 1f\n" );
1002     fprintf( outfile, "\tsubl %%ecx,%%esp\n" );
1003     fprintf( outfile, "\tleal 12(%%ebp),%%esi\n" );  /* get %esp at time of call */
1004     fprintf( outfile, "\tmovl %%esp,%%edi\n" );
1005     fprintf( outfile, "\tshrl $2,%%ecx\n" );
1006     fprintf( outfile, "\tcld\n" );
1007     fprintf( outfile, "\trep\n\tmovsl\n" );  /* copy args */
1008
1009     fprintf( outfile, "1:\tmovzbl 5(%%ebx),%%eax\n" ); /* fetch number of args to remove */
1010     fprintf( outfile, "\tleal 12(%%ebp,%%eax),%%eax\n" );
1011     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Esp) - STACK_SPACE );
1012
1013     /* Call the entry point */
1014
1015     fprintf( outfile, "\tcall *0(%%ebx)\n" );
1016     fprintf( outfile, "\tleal -%d(%%ebp),%%ecx\n", STACK_SPACE );
1017
1018     /* Restore the context structure */
1019
1020     fprintf( outfile, "2:\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegEs) );
1021     fprintf( outfile, "\tpopl %%es\n" );
1022     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegFs) );
1023     fprintf( outfile, "\tpopl %%fs\n" );
1024     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegGs) );
1025     fprintf( outfile, "\tpopl %%gs\n" );
1026
1027     fprintf( outfile, "\tmovl %d(%%ecx),%%edi\n", CONTEXTOFFSET(Edi) );
1028     fprintf( outfile, "\tmovl %d(%%ecx),%%esi\n", CONTEXTOFFSET(Esi) );
1029     fprintf( outfile, "\tmovl %d(%%ecx),%%edx\n", CONTEXTOFFSET(Edx) );
1030     fprintf( outfile, "\tmovl %d(%%ecx),%%ebx\n", CONTEXTOFFSET(Ebx) );
1031     fprintf( outfile, "\tmovl %d(%%ecx),%%eax\n", CONTEXTOFFSET(Eax) );
1032     fprintf( outfile, "\tmovl %d(%%ecx),%%ebp\n", CONTEXTOFFSET(Ebp) );
1033
1034     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegSs) );
1035     fprintf( outfile, "\tpopl %%ss\n" );
1036     fprintf( outfile, "\tmovl %d(%%ecx),%%esp\n", CONTEXTOFFSET(Esp) );
1037
1038     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(EFlags) );
1039     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegCs) );
1040     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(Eip) );
1041     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegDs) );
1042     fprintf( outfile, "\tmovl %d(%%ecx),%%ecx\n", CONTEXTOFFSET(Ecx) );
1043
1044     fprintf( outfile, "\tpopl %%ds\n" );
1045     fprintf( outfile, "\tiret\n" );
1046     function_footer( outfile, "__wine_call_from_32_regs" );
1047
1048     function_header( outfile, "__wine_call_from_32_restore_regs" );
1049     fprintf( outfile, "\tleal 4(%%esp),%%ecx\n" );
1050     fprintf( outfile, "\tjmp 2b\n" );
1051     function_footer( outfile, "__wine_call_from_32_restore_regs" );
1052 }
1053
1054
1055 /*******************************************************************
1056  *         BuildPendingEventCheck
1057  *
1058  * Build a function that checks whether there are any
1059  * pending DPMI events.
1060  *
1061  * Stack layout:
1062  *
1063  * (sp+12) long   eflags
1064  * (sp+6)  long   cs
1065  * (sp+2)  long   ip
1066  * (sp)    word   fs
1067  *
1068  * On entry to function, fs register points to a valid TEB.
1069  * On exit from function, stack will be popped.
1070  */
1071 static void BuildPendingEventCheck( FILE *outfile )
1072 {
1073     /* Function header */
1074
1075     function_header( outfile, "DPMI_PendingEventCheck" );
1076
1077     /* Check for pending events. */
1078
1079     fprintf( outfile, "\t.byte 0x64\n\ttestl $0xffffffff,(%d)\n",
1080              STRUCTOFFSET(TEB,vm86_pending) );
1081     fprintf( outfile, "\tje " __ASM_NAME("DPMI_PendingEventCheck_Cleanup") "\n" );
1082
1083     fprintf( outfile, "\t.byte 0x64\n\ttestl $0xffffffff,(%d)\n",
1084              STRUCTOFFSET(TEB,dpmi_vif) );
1085
1086     fprintf( outfile, "\tje " __ASM_NAME("DPMI_PendingEventCheck_Cleanup") "\n" );
1087
1088     /* Process pending events. */
1089
1090     fprintf( outfile, "\tsti\n" );
1091
1092     /* Start cleanup. Restore fs register. */
1093
1094     fprintf( outfile, ".globl " __ASM_NAME("DPMI_PendingEventCheck_Cleanup") "\n" );
1095     fprintf( outfile, __ASM_NAME("DPMI_PendingEventCheck_Cleanup") ":\n" );
1096     fprintf( outfile, "\tpopw %%fs\n" );
1097
1098     /* Return from function. */
1099
1100     fprintf( outfile, ".globl " __ASM_NAME("DPMI_PendingEventCheck_Return") "\n" );
1101     fprintf( outfile, __ASM_NAME("DPMI_PendingEventCheck_Return") ":\n" );
1102     fprintf( outfile, "\tiret\n" );
1103
1104     function_footer( outfile, "DPMI_PendingEventCheck" );
1105 }
1106
1107
1108 /*******************************************************************
1109  *         BuildRelays16
1110  *
1111  * Build all the 16-bit relay callbacks
1112  */
1113 void BuildRelays16( FILE *outfile )
1114 {
1115     /* File header */
1116
1117     fprintf( outfile, "/* File generated automatically. Do not edit! */\n\n" );
1118     fprintf( outfile, "\t.text\n" );
1119
1120     fprintf( outfile, __ASM_NAME("__wine_spec_thunk_text_16") ":\n\n" );
1121
1122     fprintf( outfile, __ASM_NAME("Call16_Start") ":\n" );
1123     fprintf( outfile, "\t.globl " __ASM_NAME("Call16_Start") "\n" );
1124     fprintf( outfile, "\t.byte 0\n\n" );
1125
1126     /* Standard CallFrom16 routine (WORD return) */
1127     BuildCallFrom16Core( outfile, FALSE, FALSE, TRUE );
1128
1129     /* Standard CallFrom16 routine (DWORD return) */
1130     BuildCallFrom16Core( outfile, FALSE, FALSE, FALSE );
1131
1132     /* Register CallFrom16 routine */
1133     BuildCallFrom16Core( outfile, TRUE, FALSE, FALSE );
1134
1135     /* C16ThkSL CallFrom16 routine */
1136     BuildCallFrom16Core( outfile, FALSE, TRUE, FALSE );
1137
1138     /* Standard CallTo16 routine */
1139     BuildCallTo16Core( outfile, 0 );
1140
1141     /* Register CallTo16 routine */
1142     BuildCallTo16Core( outfile, 1 );
1143
1144     /* CBClientThunkSL routine */
1145     BuildCallTo32CBClient( outfile, FALSE );
1146
1147     /* CBClientThunkSLEx routine */
1148     BuildCallTo32CBClient( outfile, TRUE  );
1149
1150     fprintf( outfile, __ASM_NAME("Call16_End") ":\n" );
1151     fprintf( outfile, "\t.globl " __ASM_NAME("Call16_End") "\n" );
1152     fprintf( outfile, "\t.size " __ASM_NAME("__wine_spec_thunk_text_16") ",. - " __ASM_NAME("__wine_spec_thunk_text_16") "\n" );
1153
1154     /* The whole Call16_Ret segment must lie within the .data section */
1155     fprintf( outfile, "\n\t.data\n" );
1156     fprintf( outfile, __ASM_NAME("__wine_spec_thunk_data_16") ":\n\n" );
1157     fprintf( outfile, "\t.globl " __ASM_NAME("Call16_Ret_Start") "\n" );
1158     fprintf( outfile, __ASM_NAME("Call16_Ret_Start") ":\n" );
1159
1160     /* Standard CallTo16 return stub */
1161     BuildRet16Func( outfile );
1162
1163     /* CBClientThunkSL return stub */
1164     BuildCallTo32CBClientRet( outfile, FALSE );
1165
1166     /* CBClientThunkSLEx return stub */
1167     BuildCallTo32CBClientRet( outfile, TRUE  );
1168
1169     /* Pending DPMI events check stub */
1170     BuildPendingEventCheck( outfile );
1171
1172     /* End of Call16_Ret segment */
1173     fprintf( outfile, "\n\t.globl " __ASM_NAME("Call16_Ret_End") "\n" );
1174     fprintf( outfile, __ASM_NAME("Call16_Ret_End") ":\n" );
1175     fprintf( outfile, "\t.size " __ASM_NAME("__wine_spec_thunk_data_16") ",. - " __ASM_NAME("__wine_spec_thunk_data_16") "\n" );
1176 }
1177
1178 /*******************************************************************
1179  *         BuildRelays32
1180  *
1181  * Build all the 32-bit relay callbacks
1182  */
1183 void BuildRelays32( FILE *outfile )
1184 {
1185     /* File header */
1186
1187     fprintf( outfile, "/* File generated automatically. Do not edit! */\n\n" );
1188     fprintf( outfile, "\t.text\n" );
1189     fprintf( outfile, __ASM_NAME("__wine_spec_thunk_text_32") ":\n\n" );
1190
1191     /* 32-bit register entry point */
1192     BuildCallFrom32Regs( outfile );
1193
1194     fprintf( outfile, "\t.size " __ASM_NAME("__wine_spec_thunk_text_32") ",. - " __ASM_NAME("__wine_spec_thunk_text_32") "\n" );
1195 }
1196
1197 #else /* __i386__ */
1198
1199 void BuildRelays16( FILE *outfile )
1200 {
1201     fprintf( outfile, "/* File not used with this architecture. Do not edit! */\n\n" );
1202 }
1203
1204 void BuildRelays32( FILE *outfile )
1205 {
1206     fprintf( outfile, "/* File not used with this architecture. Do not edit! */\n\n" );
1207 }
1208
1209 #endif  /* __i386__ */