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     const 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("CallTo16_TebSelector@GOT") "(%%ecx), %%edx\n" );
 173         fprintf( outfile, "\tmovw (%%edx), %%fs\n" );
 174     }
 175     else
 176         fprintf( outfile, "\tmovw " __ASM_NAME("CallTo16_TebSelector") ", %%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, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(SegEs) );
 405         fprintf( outfile, "\tpopl %%es\n" );
 406         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(SegFs) );
 407         fprintf( outfile, "\tpopl %%fs\n" );
 408         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(SegGs) );
 409         fprintf( outfile, "\tpopl %%gs\n" );
 410
 411         fprintf( outfile, "\tmovl %d(%%ebx), %%ebp\n", CONTEXTOFFSET(Ebp) );
 412         fprintf( outfile, "\tmovl %d(%%ebx), %%esi\n", CONTEXTOFFSET(Esi) );
 413         fprintf( outfile, "\tmovl %d(%%ebx), %%edi\n", CONTEXTOFFSET(Edi) );
 414         fprintf( outfile, "\tmovl %d(%%ebx), %%eax\n", CONTEXTOFFSET(Eax) );
 415         fprintf( outfile, "\tmovl %d(%%ebx), %%edx\n", CONTEXTOFFSET(Edx) );
 416         fprintf( outfile, "\tmovl %d(%%ebx), %%ecx\n", CONTEXTOFFSET(Ecx) );
 417         fprintf( outfile, "\tmovl %d(%%ebx), %%ebx\n", CONTEXTOFFSET(Ebx) );
 418
 419         fprintf( outfile, "\tpopl %%ds\n" );
 420         fprintf( outfile, "\tpopfl\n" );
 421         fprintf( outfile, "\tlret\n" );
 422     }
 423     else
 424     {
 425         /* Switch stack back */
 426         fprintf( outfile, "\t.byte 0x64\n\tmovw (%d), %%ss\n", STACKOFFSET+2 );
 427         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%esp\n", STACKOFFSET );
 428         fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 429
 430         /* Restore registers */
 431         fprintf( outfile, "\tpopl %%edx\n" );
 432         fprintf( outfile, "\tpopl %%ecx\n" );
 433         fprintf( outfile, "\tpopl %%ebp\n" );
 434         fprintf( outfile, "\tpopw %%ds\n" );
 435         fprintf( outfile, "\tpopw %%es\n" );
 436         fprintf( outfile, "\tpopw %%fs\n" );
 437         fprintf( outfile, "\tpopw %%gs\n" );
 438
 439         /* Prepare return value and set flags accordingly */
 440         if ( !short_ret )
 441             fprintf( outfile, "\tshldl $16, %%eax, %%edx\n" );
 442         fprintf( outfile, "\torl %%eax, %%eax\n" );
 443
 444         /* Return to return stub which will return to caller */
 445         fprintf( outfile, "\tlret $12\n" );
 446     }
 447 }
 448
 449
 450 /*******************************************************************
 451  *         BuildCallTo16Core
 452  *
 453  * This routine builds the core routines used in 32->16 thunks:
 454  *
 455  * extern DWORD WINAPI wine_call_to_16( FARPROC16 target, DWORD cbArgs, PEXCEPTION_HANDLER handler );
 456  * extern void WINAPI wine_call_to_16_regs( CONTEXT86 *context, DWORD cbArgs, PEXCEPTION_HANDLER handler );
 457  *
 458  * These routines can be called directly from 32-bit code.
 459  *
 460  * All routines expect that the 16-bit stack contents (arguments) and the
 461  * return address (segptr to CallTo16_Ret) were already set up by the
 462  * caller; nb_args must contain the number of bytes to be conserved.  The
 463  * 16-bit SS:SP will be set accordinly.
 464  *
 465  * All other registers are either taken from the CONTEXT86 structure
 466  * or else set to default values.  The target routine address is either
 467  * given directly or taken from the CONTEXT86.
 468  */
 469 static void BuildCallTo16Core( FILE *outfile, int reg_func )
 470 {
 471     const char *name = reg_func ? "wine_call_to_16_regs" : "wine_call_to_16";
 472
 473     /* Function header */
 474     function_header( outfile, name );
 475
 476     /* Function entry sequence */
 477     fprintf( outfile, "\tpushl %%ebp\n" );
 478     fprintf( outfile, "\tmovl %%esp, %%ebp\n" );
 479
 480     /* Save the 32-bit registers */
 481     fprintf( outfile, "\tpushl %%ebx\n" );
 482     fprintf( outfile, "\tpushl %%esi\n" );
 483     fprintf( outfile, "\tpushl %%edi\n" );
 484     fprintf( outfile, "\t.byte 0x64\n\tmovl %%gs,(%d)\n", STRUCTOFFSET(TEB,gs_sel) );
 485
 486     /* Setup exception frame */
 487     fprintf( outfile, "\t.byte 0x64\n\tpushl (%d)\n", STACKOFFSET );
 488     fprintf( outfile, "\tpushl 16(%%ebp)\n" ); /* handler */
 489     fprintf( outfile, "\t.byte 0x64\n\tpushl (%d)\n", STRUCTOFFSET(TEB,Tib.ExceptionList) );
 490     fprintf( outfile, "\t.byte 0x64\n\tmovl %%esp,(%d)\n", STRUCTOFFSET(TEB,Tib.ExceptionList) );
 491
 492     /* Call the actual CallTo16 routine (simulate a lcall) */
 493     fprintf( outfile, "\tpushl %%cs\n" );
 494     fprintf( outfile, "\tcall .L%s\n", name );
 495
 496     /* Remove exception frame */
 497     fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STRUCTOFFSET(TEB,Tib.ExceptionList) );
 498     fprintf( outfile, "\taddl $4, %%esp\n" );
 499     fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 500
 501     if ( !reg_func )
 502     {
 503         /* Convert return value */
 504         fprintf( outfile, "\tandl $0xffff,%%eax\n" );
 505         fprintf( outfile, "\tshll $16,%%edx\n" );
 506         fprintf( outfile, "\torl %%edx,%%eax\n" );
 507     }
 508     else
 509     {
 510         /*
 511          * Modify CONTEXT86 structure to contain new values
 512          *
 513          * NOTE:  We restore only EAX, EBX, EDX, EDX, EBP, and ESP.
 514          *        The segment registers as well as ESI and EDI should
 515          *        not be modified by a well-behaved 16-bit routine in
 516          *        any case.  [If necessary, we could restore them as well,
 517          *        at the cost of a somewhat less efficient return path.]
 518          */
 519
 520         fprintf( outfile, "\tmovl %d(%%esp), %%edi\n", STACK32OFFSET(target) - STACK32OFFSET(edi));
 521                 /* everything above edi has been popped already */
 522
 523         fprintf( outfile, "\tmovl %%eax, %d(%%edi)\n", CONTEXTOFFSET(Eax) );
 524         fprintf( outfile, "\tmovl %%ebx, %d(%%edi)\n", CONTEXTOFFSET(Ebx) );
 525         fprintf( outfile, "\tmovl %%ecx, %d(%%edi)\n", CONTEXTOFFSET(Ecx) );
 526         fprintf( outfile, "\tmovl %%edx, %d(%%edi)\n", CONTEXTOFFSET(Edx) );
 527         fprintf( outfile, "\tmovl %%ebp, %d(%%edi)\n", CONTEXTOFFSET(Ebp) );
 528         fprintf( outfile, "\tmovl %%esi, %d(%%edi)\n", CONTEXTOFFSET(Esp) );
 529                  /* The return glue code saved %esp into %esi */
 530     }
 531
 532     /* Restore the 32-bit registers */
 533     fprintf( outfile, "\tpopl %%edi\n" );
 534     fprintf( outfile, "\tpopl %%esi\n" );
 535     fprintf( outfile, "\tpopl %%ebx\n" );
 536
 537     /* Function exit sequence */
 538     fprintf( outfile, "\tpopl %%ebp\n" );
 539     fprintf( outfile, "\tret $12\n" );
 540
 541
 542     /* Start of the actual CallTo16 routine */
 543
 544     fprintf( outfile, ".L%s:\n", name );
 545
 546     /* Switch to the 16-bit stack */
 547     fprintf( outfile, "\tmovl %%esp,%%edx\n" );
 548 #ifdef __svr4__
 549     fprintf( outfile,"\tdata16\n");
 550 #endif
 551     fprintf( outfile, "\t.byte 0x64\n\tmovw (%d),%%ss\n", STACKOFFSET + 2);
 552     fprintf( outfile, "\t.byte 0x64\n\tmovw (%d),%%sp\n", STACKOFFSET );
 553     fprintf( outfile, "\t.byte 0x64\n\tmovl %%edx,(%d)\n", STACKOFFSET );
 554
 555     /* Make %bp point to the previous stackframe (built by CallFrom16) */
 556     fprintf( outfile, "\tmovzwl %%sp,%%ebp\n" );
 557     fprintf( outfile, "\tleal %d(%%ebp),%%ebp\n", STACK16OFFSET(bp) );
 558
 559     /* Add the specified offset to the new sp */
 560     fprintf( outfile, "\tsubw %d(%%edx), %%sp\n", STACK32OFFSET(nb_args) );
 561
 562     if (reg_func)
 563     {
 564         /* Push the called routine address */
 565         fprintf( outfile, "\tmovl %d(%%edx),%%edx\n", STACK32OFFSET(target) );
 566         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(SegCs) );
 567         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(Eip) );
 568
 569         /* Get the registers */
 570         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(SegDs) );
 571         fprintf( outfile, "\tpushl %d(%%edx)\n", CONTEXTOFFSET(SegEs) );
 572         fprintf( outfile, "\tpopl %%es\n" );
 573         fprintf( outfile, "\tpushl %d(%%edx)\n", CONTEXTOFFSET(SegFs) );
 574         fprintf( outfile, "\tpopl %%fs\n" );
 575         fprintf( outfile, "\tpushl %d(%%edx)\n", CONTEXTOFFSET(SegGs) );
 576         fprintf( outfile, "\tpopl %%gs\n" );
 577         fprintf( outfile, "\tmovl %d(%%edx),%%ebp\n", CONTEXTOFFSET(Ebp) );
 578         fprintf( outfile, "\tmovl %d(%%edx),%%esi\n", CONTEXTOFFSET(Esi) );
 579         fprintf( outfile, "\tmovl %d(%%edx),%%edi\n", CONTEXTOFFSET(Edi) );
 580         fprintf( outfile, "\tmovl %d(%%edx),%%eax\n", CONTEXTOFFSET(Eax) );
 581         fprintf( outfile, "\tmovl %d(%%edx),%%ebx\n", CONTEXTOFFSET(Ebx) );
 582         fprintf( outfile, "\tmovl %d(%%edx),%%ecx\n", CONTEXTOFFSET(Ecx) );
 583         fprintf( outfile, "\tmovl %d(%%edx),%%edx\n", CONTEXTOFFSET(Edx) );
 584
 585         /* Get the 16-bit ds */
 586         fprintf( outfile, "\tpopw %%ds\n" );
 587     }
 588     else  /* not a register function */
 589     {
 590         /* Push the called routine address */
 591         fprintf( outfile, "\tpushl %d(%%edx)\n", STACK32OFFSET(target) );
 592
 593         /* Set %fs and %gs to the value saved by the last CallFrom16 */
 594         fprintf( outfile, "\tpushw %d(%%ebp)\n", STACK16OFFSET(fs)-STACK16OFFSET(bp) );
 595         fprintf( outfile, "\tpopw %%fs\n" );
 596         fprintf( outfile, "\tpushw %d(%%ebp)\n", STACK16OFFSET(gs)-STACK16OFFSET(bp) );
 597         fprintf( outfile, "\tpopw %%gs\n" );
 598
 599         /* Set %ds and %es (and %ax just in case) equal to %ss */
 600         fprintf( outfile, "\tmovw %%ss,%%ax\n" );
 601         fprintf( outfile, "\tmovw %%ax,%%ds\n" );
 602         fprintf( outfile, "\tmovw %%ax,%%es\n" );
 603     }
 604
 605     /* Jump to the called routine */
 606     fprintf( outfile, "\t.byte 0x66\n" );
 607     fprintf( outfile, "\tlret\n" );
 608 }
 609
 610
 611 /*******************************************************************
 612  *         BuildRet16Func
 613  *
 614  * Build the return code for 16-bit callbacks
 615  */
 616 static void BuildRet16Func( FILE *outfile )
 617 {
 618     function_header( outfile, "CallTo16_Ret" );
 619
 620     /* Save %esp into %esi */
 621     fprintf( outfile, "\tmovl %%esp,%%esi\n" );
 622
 623     /* Restore 32-bit segment registers */
 624
 625     fprintf( outfile, "\t.byte 0x2e\n\tmovl " __ASM_NAME("CallTo16_DataSelector") "-" __ASM_NAME("Call16_Ret_Start") ",%%edi\n" );
 626 #ifdef __svr4__
 627     fprintf( outfile, "\tdata16\n");
 628 #endif
 629     fprintf( outfile, "\tmovw %%di,%%ds\n" );
 630 #ifdef __svr4__
 631     fprintf( outfile, "\tdata16\n");
 632 #endif
 633     fprintf( outfile, "\tmovw %%di,%%es\n" );
 634
 635     fprintf( outfile, "\t.byte 0x2e\n\tmovl " __ASM_NAME("CallTo16_TebSelector") "-" __ASM_NAME("Call16_Ret_Start") ",%%fs\n" );
 636
 637     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d),%%gs\n", STRUCTOFFSET(TEB,gs_sel) );
 638
 639     /* Restore the 32-bit stack */
 640
 641 #ifdef __svr4__
 642     fprintf( outfile, "\tdata16\n");
 643 #endif
 644     fprintf( outfile, "\tmovw %%di,%%ss\n" );
 645     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d),%%esp\n", STACKOFFSET );
 646
 647     /* Return to caller */
 648
 649     fprintf( outfile, "\tlret\n" );
 650
 651     /* Declare the return address and data selector variables */
 652
 653     fprintf( outfile, "\n\t.align %d\n", get_alignment(4) );
 654     fprintf( outfile, "\t.globl " __ASM_NAME("CallTo16_DataSelector") "\n" );
 655     fprintf( outfile, __ASM_NAME("CallTo16_DataSelector") ":\t.long 0\n" );
 656     fprintf( outfile, "\t.globl " __ASM_NAME("CallTo16_TebSelector") "\n" );
 657     fprintf( outfile, __ASM_NAME("CallTo16_TebSelector") ":\t.long 0\n" );
 658 }
 659
 660
 661 /*******************************************************************
 662  *         BuildCallTo32CBClient
 663  *
 664  * Call a CBClient relay stub from 32-bit code (KERNEL.620).
 665  *
 666  * Since the relay stub is itself 32-bit, this should not be a problem;
 667  * unfortunately, the relay stubs are expected to switch back to a
 668  * 16-bit stack (and 16-bit code) after completion :-(
 669  *
 670  * This would conflict with our 16- vs. 32-bit stack handling, so
 671  * we simply switch *back* to our 32-bit stack before returning to
 672  * the caller ...
 673  *
 674  * The CBClient relay stub expects to be called with the following
 675  * 16-bit stack layout, and with ebp and ebx pointing into the 16-bit
 676  * stack at the designated places:
 677  *
 678  *    ...
 679  *  (ebp+14) original arguments to the callback routine
 680  *  (ebp+10) far return address to original caller
 681  *  (ebp+6)  Thunklet target address
 682  *  (ebp+2)  Thunklet relay ID code
 683  *  (ebp)    BP (saved by CBClientGlueSL)
 684  *  (ebp-2)  SI (saved by CBClientGlueSL)
 685  *  (ebp-4)  DI (saved by CBClientGlueSL)
 686  *  (ebp-6)  DS (saved by CBClientGlueSL)
 687  *
 688  *   ...     buffer space used by the 16-bit side glue for temp copies
 689  *
 690  *  (ebx+4)  far return address to 16-bit side glue code
 691  *  (ebx)    saved 16-bit ss:sp (pointing to ebx+4)
 692  *
 693  * The 32-bit side glue code accesses both the original arguments (via ebp)
 694  * and the temporary copies prepared by the 16-bit side glue (via ebx).
 695  * After completion, the stub will load ss:sp from the buffer at ebx
 696  * and perform a far return to 16-bit code.
 697  *
 698  * To trick the relay stub into returning to us, we replace the 16-bit
 699  * return address to the glue code by a cs:ip pair pointing to our
 700  * return entry point (the original return address is saved first).
 701  * Our return stub thus called will then reload the 32-bit ss:esp and
 702  * return to 32-bit code (by using and ss:esp value that we have also
 703  * pushed onto the 16-bit stack before and a cs:eip values found at
 704  * that position on the 32-bit stack).  The ss:esp to be restored is
 705  * found relative to the 16-bit stack pointer at:
 706  *
 707  *  (ebx-4)   ss  (flat)
 708  *  (ebx-8)   sp  (32-bit stack pointer)
 709  *
 710  * The second variant of this routine, CALL32_CBClientEx, which is used
 711  * to implement KERNEL.621, has to cope with yet another problem: Here,
 712  * the 32-bit side directly returns to the caller of the CBClient thunklet,
 713  * restoring registers saved by CBClientGlueSL and cleaning up the stack.
 714  * As we have to return to our 32-bit code first, we have to adapt the
 715  * layout of our temporary area so as to include values for the registers
 716  * that are to be restored, and later (in the implementation of KERNEL.621)
 717  * we *really* restore them. The return stub restores DS, DI, SI, and BP
 718  * from the stack, skips the next 8 bytes (CBClient relay code / target),
 719  * and then performs a lret NN, where NN is the number of arguments to be
 720  * removed. Thus, we prepare our temporary area as follows:
 721  *
 722  *     (ebx+22) 16-bit cs  (this segment)
 723  *     (ebx+20) 16-bit ip  ('16-bit' return entry point)
 724  *     (ebx+16) 32-bit ss  (flat)
 725  *     (ebx+12) 32-bit sp  (32-bit stack pointer)
 726  *     (ebx+10) 16-bit bp  (points to ebx+24)
 727  *     (ebx+8)  16-bit si  (ignored)
 728  *     (ebx+6)  16-bit di  (ignored)
 729  *     (ebx+4)  16-bit ds  (we actually use the flat DS here)
 730  *     (ebx+2)  16-bit ss  (16-bit stack segment)
 731  *     (ebx+0)  16-bit sp  (points to ebx+4)
 732  *
 733  * Note that we ensure that DS is not changed and remains the flat segment,
 734  * and the 32-bit stack pointer our own return stub needs fits just
 735  * perfectly into the 8 bytes that are skipped by the Windows stub.
 736  * One problem is that we have to determine the number of removed arguments,
 737  * as these have to be really removed in KERNEL.621. Thus, the BP value
 738  * that we place in the temporary area to be restored, contains the value
 739  * that SP would have if no arguments were removed. By comparing the actual
 740  * value of SP with this value in our return stub we can compute the number
 741  * of removed arguments. This is then returned to KERNEL.621.
 742  *
 743  * The stack layout of this function:
 744  * (ebp+20)  nArgs     pointer to variable receiving nr. of args (Ex only)
 745  * (ebp+16)  esi       pointer to caller's esi value
 746  * (ebp+12)  arg       ebp value to be set for relay stub
 747  * (ebp+8)   func      CBClient relay stub address
 748  * (ebp+4)   ret addr
 749  * (ebp)     ebp
 750  */
 751 static void BuildCallTo32CBClient( FILE *outfile, BOOL isEx )
 752 {
 753     const char *name = isEx? "CBClientEx" : "CBClient";
 754     int size = isEx? 24 : 12;
 755
 756     /* Function header */
 757
 758     fprintf( outfile, "\n\t.align %d\n", get_alignment(4) );
 759     if (use_stabs)
 760         fprintf( outfile, ".stabs \"CALL32_%s:F1\",36,0,0," __ASM_NAME("CALL32_%s") "\n",
 761                  name, name );
 762     fprintf( outfile, "\t.globl " __ASM_NAME("CALL32_%s") "\n", name );
 763     fprintf( outfile, __ASM_NAME("CALL32_%s") ":\n", name );
 764
 765     /* Entry code */
 766
 767     fprintf( outfile, "\tpushl %%ebp\n" );
 768     fprintf( outfile, "\tmovl %%esp,%%ebp\n" );
 769     fprintf( outfile, "\tpushl %%edi\n" );
 770     fprintf( outfile, "\tpushl %%esi\n" );
 771     fprintf( outfile, "\tpushl %%ebx\n" );
 772
 773     /* Get the 16-bit stack */
 774
 775     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d),%%ebx\n", STACKOFFSET);
 776
 777     /* Convert it to a flat address */
 778
 779     fprintf( outfile, "\tshldl $16,%%ebx,%%eax\n" );
 780     fprintf( outfile, "\tandl $0xfff8,%%eax\n" );
 781     fprintf( outfile, "\tshrl $1,%%eax\n" );
 782     fprintf( outfile, "\tmovl " __ASM_NAME("wine_ldt_copy") "(%%eax),%%esi\n" );
 783     fprintf( outfile, "\tmovw %%bx,%%ax\n" );
 784     fprintf( outfile, "\taddl %%eax,%%esi\n" );
 785
 786     /* Allocate temporary area (simulate STACK16_PUSH) */
 787
 788     fprintf( outfile, "\tpushf\n" );
 789     fprintf( outfile, "\tcld\n" );
 790     fprintf( outfile, "\tleal -%d(%%esi), %%edi\n", size );
 791     fprintf( outfile, "\tmovl $%d, %%ecx\n", sizeof(STACK16FRAME) );
 792     fprintf( outfile, "\trep\n\tmovsb\n" );
 793     fprintf( outfile, "\tpopf\n" );
 794
 795     fprintf( outfile, "\t.byte 0x64\n\tsubw $%d,(%d)\n", size, STACKOFFSET );
 796
 797     fprintf( outfile, "\tpushl %%edi\n" );  /* remember address */
 798
 799     /* Set up temporary area */
 800
 801     if ( !isEx )
 802     {
 803         fprintf( outfile, "\tleal 4(%%edi), %%edi\n" );
 804
 805         fprintf( outfile, "\tleal -8(%%esp), %%eax\n" );
 806         fprintf( outfile, "\tmovl %%eax, -8(%%edi)\n" );    /* 32-bit sp */
 807
 808         fprintf( outfile, "\tmovw %%ss, %%ax\n" );
 809         fprintf( outfile, "\tandl $0x0000ffff, %%eax\n" );
 810         fprintf( outfile, "\tmovl %%eax, -4(%%edi)\n" );    /* 32-bit ss */
 811
 812         fprintf( outfile, "\taddl $%d, %%ebx\n", sizeof(STACK16FRAME)-size+4 + 4 );
 813         fprintf( outfile, "\tmovl %%ebx, 0(%%edi)\n" );    /* 16-bit ss:sp */
 814
 815         fprintf( outfile, "\tmovl " __ASM_NAME("CALL32_%s_RetAddr") ", %%eax\n", name );
 816         fprintf( outfile, "\tmovl %%eax, 4(%%edi)\n" );   /* overwrite return address */
 817     }
 818     else
 819     {
 820         fprintf( outfile, "\taddl $%d, %%ebx\n", sizeof(STACK16FRAME)-size+4 );
 821         fprintf( outfile, "\tmovl %%ebx, 0(%%edi)\n" );
 822
 823         fprintf( outfile, "\tmovw %%ds, %%ax\n" );
 824         fprintf( outfile, "\tmovw %%ax, 4(%%edi)\n" );
 825
 826         fprintf( outfile, "\taddl $20, %%ebx\n" );
 827         fprintf( outfile, "\tmovw %%bx, 10(%%edi)\n" );
 828
 829         fprintf( outfile, "\tleal -8(%%esp), %%eax\n" );
 830         fprintf( outfile, "\tmovl %%eax, 12(%%edi)\n" );
 831
 832         fprintf( outfile, "\tmovw %%ss, %%ax\n" );
 833         fprintf( outfile, "\tandl $0x0000ffff, %%eax\n" );
 834         fprintf( outfile, "\tmovl %%eax, 16(%%edi)\n" );
 835
 836         fprintf( outfile, "\tmovl " __ASM_NAME("CALL32_%s_RetAddr") ", %%eax\n", name );
 837         fprintf( outfile, "\tmovl %%eax, 20(%%edi)\n" );
 838     }
 839
 840     /* Set up registers and call CBClient relay stub (simulating a far call) */
 841
 842     fprintf( outfile, "\tmovl 16(%%ebp), %%esi\n" );
 843     fprintf( outfile, "\tmovl (%%esi), %%esi\n" );
 844
 845     fprintf( outfile, "\tmovl %%edi, %%ebx\n" );
 846     fprintf( outfile, "\tmovl 8(%%ebp), %%eax\n" );
 847     fprintf( outfile, "\tmovl 12(%%ebp), %%ebp\n" );
 848
 849     fprintf( outfile, "\tpushl %%cs\n" );
 850     fprintf( outfile, "\tcall *%%eax\n" );
 851
 852     /* Return new esi value to caller */
 853
 854     fprintf( outfile, "\tmovl 32(%%esp), %%edi\n" );
 855     fprintf( outfile, "\tmovl %%esi, (%%edi)\n" );
 856
 857     /* Cleanup temporary area (simulate STACK16_POP) */
 858
 859     fprintf( outfile, "\tpop %%esi\n" );
 860
 861     fprintf( outfile, "\tpushf\n" );
 862     fprintf( outfile, "\tstd\n" );
 863     fprintf( outfile, "\tdec %%esi\n" );
 864     fprintf( outfile, "\tleal %d(%%esi), %%edi\n", size );
 865     fprintf( outfile, "\tmovl $%d, %%ecx\n", sizeof(STACK16FRAME) );
 866     fprintf( outfile, "\trep\n\tmovsb\n" );
 867     fprintf( outfile, "\tpopf\n" );
 868
 869     fprintf( outfile, "\t.byte 0x64\n\taddw $%d,(%d)\n", size, STACKOFFSET );
 870
 871     /* Return argument size to caller */
 872     if ( isEx )
 873     {
 874         fprintf( outfile, "\tmovl 32(%%esp), %%ebx\n" );
 875         fprintf( outfile, "\tmovl %%ebp, (%%ebx)\n" );
 876     }
 877
 878     /* Restore registers and return */
 879
 880     fprintf( outfile, "\tpopl %%ebx\n" );
 881     fprintf( outfile, "\tpopl %%esi\n" );
 882     fprintf( outfile, "\tpopl %%edi\n" );
 883     fprintf( outfile, "\tpopl %%ebp\n" );
 884     fprintf( outfile, "\tret\n" );
 885 }
 886
 887 static void BuildCallTo32CBClientRet( FILE *outfile, BOOL isEx )
 888 {
 889     const char *name = isEx? "CBClientEx" : "CBClient";
 890
 891     /* '16-bit' return stub */
 892
 893     fprintf( outfile, "\n\t.globl " __ASM_NAME("CALL32_%s_Ret") "\n", name );
 894     fprintf( outfile, __ASM_NAME("CALL32_%s_Ret") ":\n", name );
 895
 896     if ( !isEx )
 897     {
 898         fprintf( outfile, "\tmovzwl %%sp, %%ebx\n" );
 899         fprintf( outfile, "\tlssl %%ss:-16(%%ebx), %%esp\n" );
 900     }
 901     else
 902     {
 903         fprintf( outfile, "\tmovzwl %%bp, %%ebx\n" );
 904         fprintf( outfile, "\tsubw %%bp, %%sp\n" );
 905         fprintf( outfile, "\tmovzwl %%sp, %%ebp\n" );
 906         fprintf( outfile, "\tlssl %%ss:-12(%%ebx), %%esp\n" );
 907     }
 908     fprintf( outfile, "\tlret\n" );
 909
 910     /* Declare the return address variable */
 911
 912     fprintf( outfile, "\n\t.globl " __ASM_NAME("CALL32_%s_RetAddr") "\n", name );
 913     fprintf( outfile, __ASM_NAME("CALL32_%s_RetAddr") ":\t.long 0\n", name );
 914 }
 915
 916
 917 /*******************************************************************
 918  *         BuildCallFrom32Regs
 919  *
 920  * Build a 32-bit-to-Wine call-back function for a 'register' function.
 921  * 'args' is the number of dword arguments.
 922  *
 923  * Stack layout:
 924  *   ...
 925  * (ebp+12)  first arg
 926  * (ebp+8)   ret addr to user code
 927  * (ebp+4)   ret addr to relay code
 928  * (ebp+0)   saved ebp
 929  * (ebp-128) buffer area to allow stack frame manipulation
 930  * (ebp-332) CONTEXT86 struct
 931  * (ebp-336) CONTEXT86 *argument
 932  *  ....     other arguments copied from (ebp+12)
 933  *
 934  * The entry point routine is called with a CONTEXT* extra argument,
 935  * following the normal args. In this context structure, EIP_reg
 936  * contains the return address to user code, and ESP_reg the stack
 937  * pointer on return (with the return address and arguments already
 938  * removed).
 939  */
 940 static void BuildCallFrom32Regs( FILE *outfile )
 941 {
 942     static const int STACK_SPACE = 128 + sizeof(CONTEXT86);
 943
 944     /* Function header */
 945
 946     function_header( outfile, "__wine_call_from_32_regs" );
 947
 948     /* Allocate some buffer space on the stack */
 949
 950     fprintf( outfile, "\tpushl %%ebp\n" );
 951     fprintf( outfile, "\tmovl %%esp,%%ebp\n ");
 952     fprintf( outfile, "\tleal -%d(%%esp), %%esp\n", STACK_SPACE );
 953
 954     /* Build the context structure */
 955
 956     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Eax) - STACK_SPACE );
 957     fprintf( outfile, "\tpushfl\n" );
 958     fprintf( outfile, "\tpopl %%eax\n" );
 959     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(EFlags) - STACK_SPACE );
 960     fprintf( outfile, "\tmovl 0(%%ebp),%%eax\n" );
 961     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Ebp) - STACK_SPACE );
 962     fprintf( outfile, "\tmovl %%ebx,%d(%%ebp)\n", CONTEXTOFFSET(Ebx) - STACK_SPACE );
 963     fprintf( outfile, "\tmovl %%ecx,%d(%%ebp)\n", CONTEXTOFFSET(Ecx) - STACK_SPACE );
 964     fprintf( outfile, "\tmovl %%edx,%d(%%ebp)\n", CONTEXTOFFSET(Edx) - STACK_SPACE );
 965     fprintf( outfile, "\tmovl %%esi,%d(%%ebp)\n", CONTEXTOFFSET(Esi) - STACK_SPACE );
 966     fprintf( outfile, "\tmovl %%edi,%d(%%ebp)\n", CONTEXTOFFSET(Edi) - STACK_SPACE );
 967
 968     fprintf( outfile, "\txorl %%eax,%%eax\n" );
 969     fprintf( outfile, "\tmovw %%cs,%%ax\n" );
 970     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegCs) - STACK_SPACE );
 971     fprintf( outfile, "\tmovw %%es,%%ax\n" );
 972     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegEs) - STACK_SPACE );
 973     fprintf( outfile, "\tmovw %%fs,%%ax\n" );
 974     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegFs) - STACK_SPACE );
 975     fprintf( outfile, "\tmovw %%gs,%%ax\n" );
 976     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegGs) - STACK_SPACE );
 977     fprintf( outfile, "\tmovw %%ss,%%ax\n" );
 978     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegSs) - STACK_SPACE );
 979     fprintf( outfile, "\tmovw %%ds,%%ax\n" );
 980     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegDs) - STACK_SPACE );
 981     fprintf( outfile, "\tmovw %%ax,%%es\n" );  /* set %es equal to %ds just in case */
 982
 983     fprintf( outfile, "\tmovl $0x%x,%%eax\n", CONTEXT86_FULL );
 984     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(ContextFlags) - STACK_SPACE );
 985
 986     fprintf( outfile, "\tmovl 8(%%ebp),%%eax\n" ); /* Get %eip at time of call */
 987     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Eip) - STACK_SPACE );
 988
 989     /* Transfer the arguments */
 990
 991     fprintf( outfile, "\tmovl 4(%%ebp),%%ebx\n" );   /* get relay code addr */
 992     fprintf( outfile, "\tpushl %%esp\n" );           /* push ptr to context struct */
 993     fprintf( outfile, "\tmovzbl 4(%%ebx),%%ecx\n" ); /* fetch number of args to copy */
 994     fprintf( outfile, "\tjecxz 1f\n" );
 995     fprintf( outfile, "\tsubl %%ecx,%%esp\n" );
 996     fprintf( outfile, "\tleal 12(%%ebp),%%esi\n" );  /* get %esp at time of call */
 997     fprintf( outfile, "\tmovl %%esp,%%edi\n" );
 998     fprintf( outfile, "\tshrl $2,%%ecx\n" );
 999     fprintf( outfile, "\tcld\n" );
1000     fprintf( outfile, "\trep\n\tmovsl\n" );  /* copy args */
1001
1002     fprintf( outfile, "1:\tmovzbl 5(%%ebx),%%eax\n" ); /* fetch number of args to remove */
1003     fprintf( outfile, "\tleal 12(%%ebp,%%eax),%%eax\n" );
1004     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Esp) - STACK_SPACE );
1005
1006     /* Call the entry point */
1007
1008     fprintf( outfile, "\tcall *0(%%ebx)\n" );
1009     fprintf( outfile, "\tleal -%d(%%ebp),%%ecx\n", STACK_SPACE );
1010
1011     /* Restore the context structure */
1012
1013     fprintf( outfile, "2:\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegEs) );
1014     fprintf( outfile, "\tpopl %%es\n" );
1015     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegFs) );
1016     fprintf( outfile, "\tpopl %%fs\n" );
1017     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegGs) );
1018     fprintf( outfile, "\tpopl %%gs\n" );
1019
1020     fprintf( outfile, "\tmovl %d(%%ecx),%%edi\n", CONTEXTOFFSET(Edi) );
1021     fprintf( outfile, "\tmovl %d(%%ecx),%%esi\n", CONTEXTOFFSET(Esi) );
1022     fprintf( outfile, "\tmovl %d(%%ecx),%%edx\n", CONTEXTOFFSET(Edx) );
1023     fprintf( outfile, "\tmovl %d(%%ecx),%%ebx\n", CONTEXTOFFSET(Ebx) );
1024     fprintf( outfile, "\tmovl %d(%%ecx),%%eax\n", CONTEXTOFFSET(Eax) );
1025     fprintf( outfile, "\tmovl %d(%%ecx),%%ebp\n", CONTEXTOFFSET(Ebp) );
1026
1027     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegSs) );
1028     fprintf( outfile, "\tpopl %%ss\n" );
1029     fprintf( outfile, "\tmovl %d(%%ecx),%%esp\n", CONTEXTOFFSET(Esp) );
1030
1031     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(EFlags) );
1032     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegCs) );
1033     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(Eip) );
1034     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegDs) );
1035     fprintf( outfile, "\tmovl %d(%%ecx),%%ecx\n", CONTEXTOFFSET(Ecx) );
1036
1037     fprintf( outfile, "\tpopl %%ds\n" );
1038     fprintf( outfile, "\tiret\n" );
1039
1040     function_header( outfile, "__wine_call_from_32_restore_regs" );
1041     fprintf( outfile, "\tleal 4(%%esp),%%ecx\n" );
1042     fprintf( outfile, "\tjmp 2b\n" );
1043 }
1044
1045
1046 /*******************************************************************
1047  *         BuildPendingEventCheck
1048  *
1049  * Build a function that checks whether there are any
1050  * pending DPMI events.
1051  *
1052  * Stack layout:
1053  *
1054  * (sp+12) long   eflags
1055  * (sp+6)  long   cs
1056  * (sp+2)  long   ip
1057  * (sp)    word   fs
1058  *
1059  * On entry to function, fs register points to a valid TEB.
1060  * On exit from function, stack will be popped.
1061  */
1062 void BuildPendingEventCheck( FILE *outfile )
1063 {
1064     /* Function header */
1065
1066     function_header( outfile, "DPMI_PendingEventCheck" );
1067
1068     /* Check for pending events. */
1069
1070     fprintf( outfile, "\t.byte 0x64\n\ttestl $0xffffffff,(%d)\n",
1071              STRUCTOFFSET(TEB,vm86_pending) );
1072     fprintf( outfile, "\tje " __ASM_NAME("DPMI_PendingEventCheck_Cleanup") "\n" );
1073
1074     fprintf( outfile, "\t.byte 0x64\n\ttestl $0xffffffff,(%d)\n",
1075              STRUCTOFFSET(TEB,dpmi_vif) );
1076
1077     fprintf( outfile, "\tje " __ASM_NAME("DPMI_PendingEventCheck_Cleanup") "\n" );
1078
1079     /* Process pending events. */
1080
1081     fprintf( outfile, "\tsti\n" );
1082
1083     /* Start cleanup. Restore fs register. */
1084
1085     fprintf( outfile, ".globl " __ASM_NAME("DPMI_PendingEventCheck_Cleanup") "\n" );
1086     fprintf( outfile, __ASM_NAME("DPMI_PendingEventCheck_Cleanup") ":\n" );
1087     fprintf( outfile, "\tpopw %%fs\n" );
1088
1089     /* Return from function. */
1090
1091     fprintf( outfile, ".globl " __ASM_NAME("DPMI_PendingEventCheck_Return") "\n" );
1092     fprintf( outfile, __ASM_NAME("DPMI_PendingEventCheck_Return") ":\n" );
1093     fprintf( outfile, "\tiret\n" );
1094 }
1095
1096
1097 /*******************************************************************
1098  *         BuildRelays16
1099  *
1100  * Build all the 16-bit relay callbacks
1101  */
1102 void BuildRelays16( FILE *outfile )
1103 {
1104     char buffer[1024];
1105
1106     /* File header */
1107
1108     fprintf( outfile, "/* File generated automatically. Do not edit! */\n\n" );
1109     fprintf( outfile, "\t.text\n" );
1110
1111     if (output_file_name && use_stabs)
1112     {
1113         getcwd(buffer, sizeof(buffer));
1114         fprintf( outfile, "\t.file\t\"%s\"\n", output_file_name );
1115
1116         /*
1117          * The stabs help the internal debugger as they are an indication that it
1118          * is sensible to step into a thunk/trampoline.
1119          */
1120         fprintf( outfile, ".stabs \"%s/\",100,0,0,Code_Start\n", buffer);
1121         fprintf( outfile, ".stabs \"%s\",100,0,0,Code_Start\n", output_file_name );
1122         fprintf( outfile, "Code_Start:\n\n" );
1123     }
1124
1125     fprintf( outfile, __ASM_NAME("Call16_Start") ":\n" );
1126     fprintf( outfile, "\t.globl " __ASM_NAME("Call16_Start") "\n" );
1127     fprintf( outfile, "\t.byte 0\n\n" );
1128
1129     /* Standard CallFrom16 routine (WORD return) */
1130     BuildCallFrom16Core( outfile, FALSE, FALSE, TRUE );
1131
1132     /* Standard CallFrom16 routine (DWORD return) */
1133     BuildCallFrom16Core( outfile, FALSE, FALSE, FALSE );
1134
1135     /* Register CallFrom16 routine */
1136     BuildCallFrom16Core( outfile, TRUE, FALSE, FALSE );
1137
1138     /* C16ThkSL CallFrom16 routine */
1139     BuildCallFrom16Core( outfile, FALSE, TRUE, FALSE );
1140
1141     /* Standard CallTo16 routine */
1142     BuildCallTo16Core( outfile, 0 );
1143
1144     /* Register CallTo16 routine */
1145     BuildCallTo16Core( outfile, 1 );
1146
1147     /* CBClientThunkSL routine */
1148     BuildCallTo32CBClient( outfile, FALSE );
1149
1150     /* CBClientThunkSLEx routine */
1151     BuildCallTo32CBClient( outfile, TRUE  );
1152
1153     fprintf( outfile, __ASM_NAME("Call16_End") ":\n" );
1154     fprintf( outfile, "\t.globl " __ASM_NAME("Call16_End") "\n" );
1155
1156     if (use_stabs)
1157     {
1158         fprintf( outfile, "\t.stabs \"\",100,0,0,.Letext\n");
1159         fprintf( outfile, ".Letext:\n");
1160
1161         /* Some versions of gdb need to have the filename data for
1162            each section, so output it again for the data section. */
1163         if (output_file_name)
1164         {
1165             fprintf( outfile, ".stabs \"%s/\",100,0,0,Data_Start\n", buffer);
1166             fprintf( outfile, ".stabs \"%s\",100,0,0,Data_Start\n", output_file_name );
1167             fprintf( outfile, "Data_Start:\n\n" );
1168         }
1169     }
1170
1171     /* The whole Call16_Ret segment must lie within the .data section */
1172     fprintf( outfile, "\n\t.data\n" );
1173     fprintf( outfile, "\t.globl " __ASM_NAME("Call16_Ret_Start") "\n" );
1174     fprintf( outfile, __ASM_NAME("Call16_Ret_Start") ":\n" );
1175
1176     /* Standard CallTo16 return stub */
1177     BuildRet16Func( outfile );
1178
1179     /* CBClientThunkSL return stub */
1180     BuildCallTo32CBClientRet( outfile, FALSE );
1181
1182     /* CBClientThunkSLEx return stub */
1183     BuildCallTo32CBClientRet( outfile, TRUE  );
1184
1185     /* Pending DPMI events check stub */
1186     BuildPendingEventCheck( outfile );
1187
1188     /* End of Call16_Ret segment */
1189     fprintf( outfile, "\n\t.globl " __ASM_NAME("Call16_Ret_End") "\n" );
1190     fprintf( outfile, __ASM_NAME("Call16_Ret_End") ":\n" );
1191 }
1192
1193 /*******************************************************************
1194  *         BuildRelays32
1195  *
1196  * Build all the 32-bit relay callbacks
1197  */
1198 void BuildRelays32( FILE *outfile )
1199 {
1200     /* File header */
1201
1202     fprintf( outfile, "/* File generated automatically. Do not edit! */\n\n" );
1203     fprintf( outfile, "\t.text\n" );
1204
1205     if (output_file_name && use_stabs)
1206     {
1207         char buffer[1024];
1208         getcwd(buffer, sizeof(buffer));
1209         fprintf( outfile, "\t.file\t\"%s\"\n", output_file_name );
1210
1211         /*
1212          * The stabs help the internal debugger as they are an indication that it
1213          * is sensible to step into a thunk/trampoline.
1214          */
1215         fprintf( outfile, ".stabs \"%s/\",100,0,0,Code_Start\n", buffer);
1216         fprintf( outfile, ".stabs \"%s\",100,0,0,Code_Start\n", output_file_name );
1217         fprintf( outfile, "Code_Start:\n\n" );
1218     }
1219
1220     /* 32-bit register entry point */
1221     BuildCallFrom32Regs( outfile );
1222
1223     if (use_stabs)
1224     {
1225         fprintf( outfile, "\t.stabs \"\",100,0,0,.Letext\n");
1226         fprintf( outfile, ".Letext:\n");
1227     }
1228 }
1229
1230 #else /* __i386__ */
1231
1232 void BuildRelays16( FILE *outfile )
1233 {
1234     fprintf( outfile, "/* File not used with this architecture. Do not edit! */\n\n" );
1235 }
1236
1237 void BuildRelays32( FILE *outfile )
1238 {
1239     fprintf( outfile, "/* File not used with this architecture. Do not edit! */\n\n" );
1240 }
1241
1242 #endif  /* __i386__ */