reactos/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
  27 #include <ctype.h>
  28
  29 #include "thread.h"
  30 #include "wine/winbase16.h"
  31
  32 #include "build.h"
  33
  34 static void function_header( FILE *outfile, const char *name )
  35 {
  36     fprintf( outfile, "\n\t.align %d\n", get_alignment(4) );
  37     fprintf( outfile, "\t%s\n", func_declaration(name) );
  38     fprintf( outfile, "\t.globl %s\n", asm_name(name) );
  39     fprintf( outfile, "%s:\n", asm_name(name) );
  40 }
  41
  42
  43 static void function_footer( FILE *outfile, const char *name )
  44 {
  45     const char *size = func_size( name );
  46     if (size[0]) fprintf( outfile, "\t%s\n", size );
  47 }
  48
  49 static inline const char *data16_prefix(void)
  50 {
  51     return (target_platform == PLATFORM_SVR4) ? "\tdata16\n" : "";
  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 %s(%%ecx), %%edx\n", asm_name("CallTo16_DataSelector@GOT") );
 155         fprintf( outfile, "\t.byte 0x2e\n\tmovl (%%edx), %%edx\n" );
 156     }
 157     else
 158         fprintf( outfile, "\t.byte 0x2e\n\tmovl %s,%%edx\n", asm_name("CallTo16_DataSelector") );
 159
 160     /* Load 32-bit segment registers */
 161     fprintf( outfile, "%s\tmovw %%dx, %%ds\n", data16_prefix() );
 162     fprintf( outfile, "%s\tmovw %%dx, %%es\n", data16_prefix() );
 163
 164     if ( UsePIC )
 165     {
 166         fprintf( outfile, "\tmovl %s(%%ecx), %%edx\n", asm_name("CallTo16_TebSelector@GOT") );
 167         fprintf( outfile, "\tmovw (%%edx), %%fs\n" );
 168     }
 169     else
 170         fprintf( outfile, "\tmovw %s, %%fs\n", asm_name("CallTo16_TebSelector") );
 171
 172     fprintf( outfile, "\t.byte 0x64\n\tmov (%d),%%gs\n", STRUCTOFFSET(TEB,gs_sel) );
 173
 174     /* Get address of wine_ldt_copy array into %ecx */
 175     if ( UsePIC )
 176         fprintf( outfile, "\tmovl %s(%%ecx), %%ecx\n", asm_name("wine_ldt_copy@GOT") );
 177     else
 178         fprintf( outfile, "\tmovl $%s, %%ecx\n", asm_name("wine_ldt_copy") );
 179
 180     /* Translate STACK16FRAME base to flat offset in %edx */
 181     fprintf( outfile, "\tmovw %%ss, %%dx\n" );
 182     fprintf( outfile, "\tandl $0xfff8, %%edx\n" );
 183     fprintf( outfile, "\tshrl $1, %%edx\n" );
 184     fprintf( outfile, "\tmovl (%%ecx,%%edx), %%edx\n" );
 185     fprintf( outfile, "\tmovzwl %%sp, %%ebp\n" );
 186     fprintf( outfile, "\tleal (%%ebp,%%edx), %%edx\n" );
 187
 188     /* Get saved flags into %ecx */
 189     fprintf( outfile, "\tpopl %%ecx\n" );
 190
 191     /* Get the 32-bit stack pointer from the TEB and complete STACK16FRAME */
 192     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d), %%ebp\n", STACKOFFSET );
 193     fprintf( outfile, "\tpushl %%ebp\n" );
 194
 195     /* Switch stacks */
 196     fprintf( outfile, "%s\t.byte 0x64\n\tmovw %%ss, (%d)\n", data16_prefix(), STACKOFFSET + 2 );
 197     fprintf( outfile, "\t.byte 0x64\n\tmovw %%sp, (%d)\n", STACKOFFSET );
 198     fprintf( outfile, "\tpushl %%ds\n" );
 199     fprintf( outfile, "\tpopl %%ss\n" );
 200     fprintf( outfile, "\tmovl %%ebp, %%esp\n" );
 201     fprintf( outfile, "\taddl $%d, %%ebp\n", STRUCTOFFSET(STACK32FRAME, ebp) );
 202
 203
 204     /* At this point:
 205        STACK16FRAME is completely set up
 206        DS, ES, SS: flat data segment
 207        FS: current TEB
 208        ESP: points to last STACK32FRAME
 209        EBP: points to ebp member of last STACK32FRAME
 210        EDX: points to current STACK16FRAME
 211        ECX: contains saved flags
 212        all other registers: unchanged */
 213
 214     /* Special case: C16ThkSL stub */
 215     if ( thunk )
 216     {
 217         /* Set up registers as expected and call thunk */
 218         fprintf( outfile, "\tleal %d(%%edx), %%ebx\n", sizeof(STACK16FRAME)-22 );
 219         fprintf( outfile, "\tleal -4(%%esp), %%ebp\n" );
 220
 221         fprintf( outfile, "\tcall *%d(%%edx)\n", STACK16OFFSET(entry_point) );
 222
 223         /* Switch stack back */
 224         fprintf( outfile, "\t.byte 0x64\n\tmovw (%d), %%ss\n", STACKOFFSET+2 );
 225         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%esp\n", STACKOFFSET );
 226         fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 227
 228         /* Restore registers and return directly to caller */
 229         fprintf( outfile, "\taddl $8, %%esp\n" );
 230         fprintf( outfile, "\tpopl %%ebp\n" );
 231         fprintf( outfile, "\tpopw %%ds\n" );
 232         fprintf( outfile, "\tpopw %%es\n" );
 233         fprintf( outfile, "\tpopw %%fs\n" );
 234         fprintf( outfile, "\tpopw %%gs\n" );
 235         fprintf( outfile, "\taddl $20, %%esp\n" );
 236
 237         fprintf( outfile, "\txorb %%ch, %%ch\n" );
 238         fprintf( outfile, "\tpopl %%ebx\n" );
 239         fprintf( outfile, "\taddw %%cx, %%sp\n" );
 240         fprintf( outfile, "\tpush %%ebx\n" );
 241
 242         fprintf( outfile, "\t.byte 0x66\n" );
 243         fprintf( outfile, "\tlret\n" );
 244
 245         return;
 246     }
 247
 248
 249     /* Build register CONTEXT */
 250     if ( reg_func )
 251     {
 252         fprintf( outfile, "\tsubl $%d, %%esp\n", sizeof(CONTEXT86) );
 253
 254         fprintf( outfile, "\tmovl %%ecx, %d(%%esp)\n", CONTEXTOFFSET(EFlags) );
 255
 256         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Eax) );
 257         fprintf( outfile, "\tmovl %%ebx, %d(%%esp)\n", CONTEXTOFFSET(Ebx) );
 258         fprintf( outfile, "\tmovl %%esi, %d(%%esp)\n", CONTEXTOFFSET(Esi) );
 259         fprintf( outfile, "\tmovl %%edi, %d(%%esp)\n", CONTEXTOFFSET(Edi) );
 260
 261         fprintf( outfile, "\tmovl %d(%%edx), %%eax\n", STACK16OFFSET(ebp) );
 262         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Ebp) );
 263         fprintf( outfile, "\tmovl %d(%%edx), %%eax\n", STACK16OFFSET(ecx) );
 264         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Ecx) );
 265         fprintf( outfile, "\tmovl %d(%%edx), %%eax\n", STACK16OFFSET(edx) );
 266         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Edx) );
 267
 268         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(ds) );
 269         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegDs) );
 270         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(es) );
 271         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegEs) );
 272         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(fs) );
 273         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegFs) );
 274         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(gs) );
 275         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegGs) );
 276
 277         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(cs) );
 278         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegCs) );
 279         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(ip) );
 280         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Eip) );
 281
 282         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%eax\n", STACKOFFSET+2 );
 283         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegSs) );
 284         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%eax\n", STACKOFFSET );
 285         fprintf( outfile, "\taddl $%d, %%eax\n", STACK16OFFSET(ip) );
 286         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Esp) );
 287 #if 0
 288         fprintf( outfile, "\tfsave %d(%%esp)\n", CONTEXTOFFSET(FloatSave) );
 289 #endif
 290
 291         /* Push address of CONTEXT86 structure -- popped by the relay routine */
 292         fprintf( outfile, "\tpushl %%esp\n" );
 293     }
 294
 295
 296     /* Print debug info before call */
 297     if ( debugging )
 298     {
 299         if ( UsePIC )
 300         {
 301             fprintf( outfile, "\tpushl %%ebx\n" );
 302
 303             /* Get Global Offset Table into %ebx (for PLT call) */
 304             fprintf( outfile, "\tcall .L__wine_call_from_16_%s.getgot2\n", name );
 305             fprintf( outfile, ".L__wine_call_from_16_%s.getgot2:\n", name );
 306             fprintf( outfile, "\tpopl %%ebx\n" );
 307             fprintf( outfile, "\taddl $_GLOBAL_OFFSET_TABLE_+[.-.L__wine_call_from_16_%s.getgot2], %%ebx\n", name );
 308         }
 309
 310         fprintf( outfile, "\tpushl %%edx\n" );
 311         if ( reg_func )
 312             fprintf( outfile, "\tleal -%d(%%ebp), %%eax\n\tpushl %%eax\n",
 313                               sizeof(CONTEXT) + STRUCTOFFSET(STACK32FRAME, ebp) );
 314         else
 315             fprintf( outfile, "\tpushl $0\n" );
 316
 317         if ( UsePIC )
 318             fprintf( outfile, "\tcall %s\n ", asm_name("RELAY_DebugCallFrom16@PLT"));
 319         else
 320             fprintf( outfile, "\tcall %s\n ", asm_name("RELAY_DebugCallFrom16"));
 321
 322         fprintf( outfile, "\tpopl %%edx\n" );
 323         fprintf( outfile, "\tpopl %%edx\n" );
 324
 325         if ( UsePIC )
 326             fprintf( outfile, "\tpopl %%ebx\n" );
 327     }
 328
 329     /* Call relay routine (which will call the API entry point) */
 330     fprintf( outfile, "\tleal %d(%%edx), %%eax\n", sizeof(STACK16FRAME) );
 331     fprintf( outfile, "\tpushl %%eax\n" );
 332     fprintf( outfile, "\tpushl %d(%%edx)\n", STACK16OFFSET(entry_point) );
 333     fprintf( outfile, "\tcall *%d(%%edx)\n", STACK16OFFSET(relay) );
 334
 335     /* Print debug info after call */
 336     if ( debugging )
 337     {
 338         if ( UsePIC )
 339         {
 340             fprintf( outfile, "\tpushl %%ebx\n" );
 341
 342             /* Get Global Offset Table into %ebx (for PLT call) */
 343             fprintf( outfile, "\tcall .L__wine_call_from_16_%s.getgot3\n", name );
 344             fprintf( outfile, ".L__wine_call_from_16_%s.getgot3:\n", name );
 345             fprintf( outfile, "\tpopl %%ebx\n" );
 346             fprintf( outfile, "\taddl $_GLOBAL_OFFSET_TABLE_+[.-.L__wine_call_from_16_%s.getgot3], %%ebx\n", name );
 347         }
 348
 349         fprintf( outfile, "\tpushl %%eax\n" );
 350         if ( reg_func )
 351             fprintf( outfile, "\tleal -%d(%%ebp), %%eax\n\tpushl %%eax\n",
 352                               sizeof(CONTEXT) + STRUCTOFFSET(STACK32FRAME, ebp) );
 353         else
 354             fprintf( outfile, "\tpushl $0\n" );
 355
 356         if ( UsePIC )
 357             fprintf( outfile, "\tcall %s\n ", asm_name("RELAY_DebugCallFrom16Ret@PLT"));
 358         else
 359             fprintf( outfile, "\tcall %s\n ", asm_name("RELAY_DebugCallFrom16Ret"));
 360
 361         fprintf( outfile, "\tpopl %%eax\n" );
 362         fprintf( outfile, "\tpopl %%eax\n" );
 363
 364         if ( UsePIC )
 365             fprintf( outfile, "\tpopl %%ebx\n" );
 366     }
 367
 368
 369     if ( reg_func )
 370     {
 371         fprintf( outfile, "\tmovl %%esp, %%ebx\n" );
 372
 373         /* Switch stack back */
 374         fprintf( outfile, "\t.byte 0x64\n\tmovw (%d), %%ss\n", STACKOFFSET+2 );
 375         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%esp\n", STACKOFFSET );
 376         fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 377
 378         /* Get return address to CallFrom16 stub */
 379         fprintf( outfile, "\taddw $%d, %%sp\n", STACK16OFFSET(callfrom_ip)-4 );
 380         fprintf( outfile, "\tpopl %%eax\n" );
 381         fprintf( outfile, "\tpopl %%edx\n" );
 382
 383         /* Restore all registers from CONTEXT */
 384         fprintf( outfile, "\tmovw %d(%%ebx), %%ss\n", CONTEXTOFFSET(SegSs) );
 385         fprintf( outfile, "\tmovl %d(%%ebx), %%esp\n", CONTEXTOFFSET(Esp) );
 386         fprintf( outfile, "\taddl $4, %%esp\n" );  /* room for final return address */
 387
 388         fprintf( outfile, "\tpushw %d(%%ebx)\n", CONTEXTOFFSET(SegCs) );
 389         fprintf( outfile, "\tpushw %d(%%ebx)\n", CONTEXTOFFSET(Eip) );
 390         fprintf( outfile, "\tpushl %%edx\n" );
 391         fprintf( outfile, "\tpushl %%eax\n" );
 392         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(EFlags) );
 393         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(SegDs) );
 394
 395         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(SegEs) );
 396         fprintf( outfile, "\tpopl %%es\n" );
 397         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(SegFs) );
 398         fprintf( outfile, "\tpopl %%fs\n" );
 399         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(SegGs) );
 400         fprintf( outfile, "\tpopl %%gs\n" );
 401
 402         fprintf( outfile, "\tmovl %d(%%ebx), %%ebp\n", CONTEXTOFFSET(Ebp) );
 403         fprintf( outfile, "\tmovl %d(%%ebx), %%esi\n", CONTEXTOFFSET(Esi) );
 404         fprintf( outfile, "\tmovl %d(%%ebx), %%edi\n", CONTEXTOFFSET(Edi) );
 405         fprintf( outfile, "\tmovl %d(%%ebx), %%eax\n", CONTEXTOFFSET(Eax) );
 406         fprintf( outfile, "\tmovl %d(%%ebx), %%edx\n", CONTEXTOFFSET(Edx) );
 407         fprintf( outfile, "\tmovl %d(%%ebx), %%ecx\n", CONTEXTOFFSET(Ecx) );
 408         fprintf( outfile, "\tmovl %d(%%ebx), %%ebx\n", CONTEXTOFFSET(Ebx) );
 409
 410         fprintf( outfile, "\tpopl %%ds\n" );
 411         fprintf( outfile, "\tpopfl\n" );
 412         fprintf( outfile, "\tlret\n" );
 413     }
 414     else
 415     {
 416         /* Switch stack back */
 417         fprintf( outfile, "\t.byte 0x64\n\tmovw (%d), %%ss\n", STACKOFFSET+2 );
 418         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%esp\n", STACKOFFSET );
 419         fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 420
 421         /* Restore registers */
 422         fprintf( outfile, "\tpopl %%edx\n" );
 423         fprintf( outfile, "\tpopl %%ecx\n" );
 424         fprintf( outfile, "\tpopl %%ebp\n" );
 425         fprintf( outfile, "\tpopw %%ds\n" );
 426         fprintf( outfile, "\tpopw %%es\n" );
 427         fprintf( outfile, "\tpopw %%fs\n" );
 428         fprintf( outfile, "\tpopw %%gs\n" );
 429
 430         /* Prepare return value and set flags accordingly */
 431         if ( !short_ret )
 432             fprintf( outfile, "\tshldl $16, %%eax, %%edx\n" );
 433         fprintf( outfile, "\torl %%eax, %%eax\n" );
 434
 435         /* Return to return stub which will return to caller */
 436         fprintf( outfile, "\tlret $12\n" );
 437     }
 438     if (thunk) function_footer( outfile, "__wine_call_from_16_thunk" );
 439     else if (reg_func) function_footer( outfile, "__wine_call_from_16_regs" );
 440     else if (short_ret) function_footer( outfile, "__wine_call_from_16_word" );
 441     else function_footer( outfile, "__wine_call_from_16_long" );
 442 }
 443
 444
 445 /*******************************************************************
 446  *         BuildCallTo16Core
 447  *
 448  * This routine builds the core routines used in 32->16 thunks:
 449  *
 450  * extern DWORD WINAPI wine_call_to_16( FARPROC16 target, DWORD cbArgs, PEXCEPTION_HANDLER handler );
 451  * extern void WINAPI wine_call_to_16_regs( CONTEXT86 *context, DWORD cbArgs, PEXCEPTION_HANDLER handler );
 452  *
 453  * These routines can be called directly from 32-bit code.
 454  *
 455  * All routines expect that the 16-bit stack contents (arguments) and the
 456  * return address (segptr to CallTo16_Ret) were already set up by the
 457  * caller; nb_args must contain the number of bytes to be conserved.  The
 458  * 16-bit SS:SP will be set accordinly.
 459  *
 460  * All other registers are either taken from the CONTEXT86 structure
 461  * or else set to default values.  The target routine address is either
 462  * given directly or taken from the CONTEXT86.
 463  */
 464 static void BuildCallTo16Core( FILE *outfile, int reg_func )
 465 {
 466     const char *name = reg_func ? "wine_call_to_16_regs" : "wine_call_to_16";
 467
 468     /* Function header */
 469     function_header( outfile, name );
 470
 471     /* Function entry sequence */
 472     fprintf( outfile, "\tpushl %%ebp\n" );
 473     fprintf( outfile, "\tmovl %%esp, %%ebp\n" );
 474
 475     /* Save the 32-bit registers */
 476     fprintf( outfile, "\tpushl %%ebx\n" );
 477     fprintf( outfile, "\tpushl %%esi\n" );
 478     fprintf( outfile, "\tpushl %%edi\n" );
 479     fprintf( outfile, "\t.byte 0x64\n\tmov %%gs,(%d)\n", STRUCTOFFSET(TEB,gs_sel) );
 480
 481     /* Setup exception frame */
 482     fprintf( outfile, "\t.byte 0x64\n\tpushl (%d)\n", STACKOFFSET );
 483     fprintf( outfile, "\tpushl 16(%%ebp)\n" ); /* handler */
 484     fprintf( outfile, "\t.byte 0x64\n\tpushl (%d)\n", STRUCTOFFSET(TEB,Tib.ExceptionList) );
 485     fprintf( outfile, "\t.byte 0x64\n\tmovl %%esp,(%d)\n", STRUCTOFFSET(TEB,Tib.ExceptionList) );
 486
 487     /* Call the actual CallTo16 routine (simulate a lcall) */
 488     fprintf( outfile, "\tpushl %%cs\n" );
 489     fprintf( outfile, "\tcall .L%s\n", name );
 490
 491     /* Remove exception frame */
 492     fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STRUCTOFFSET(TEB,Tib.ExceptionList) );
 493     fprintf( outfile, "\taddl $4, %%esp\n" );
 494     fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 495
 496     if ( !reg_func )
 497     {
 498         /* Convert return value */
 499         fprintf( outfile, "\tandl $0xffff,%%eax\n" );
 500         fprintf( outfile, "\tshll $16,%%edx\n" );
 501         fprintf( outfile, "\torl %%edx,%%eax\n" );
 502     }
 503     else
 504     {
 505         /*
 506          * Modify CONTEXT86 structure to contain new values
 507          *
 508          * NOTE:  We restore only EAX, EBX, EDX, EDX, EBP, and ESP.
 509          *        The segment registers as well as ESI and EDI should
 510          *        not be modified by a well-behaved 16-bit routine in
 511          *        any case.  [If necessary, we could restore them as well,
 512          *        at the cost of a somewhat less efficient return path.]
 513          */
 514
 515         fprintf( outfile, "\tmovl %d(%%esp), %%edi\n", STACK32OFFSET(target) - STACK32OFFSET(edi));
 516                 /* everything above edi has been popped already */
 517
 518         fprintf( outfile, "\tmovl %%eax, %d(%%edi)\n", CONTEXTOFFSET(Eax) );
 519         fprintf( outfile, "\tmovl %%ebx, %d(%%edi)\n", CONTEXTOFFSET(Ebx) );
 520         fprintf( outfile, "\tmovl %%ecx, %d(%%edi)\n", CONTEXTOFFSET(Ecx) );
 521         fprintf( outfile, "\tmovl %%edx, %d(%%edi)\n", CONTEXTOFFSET(Edx) );
 522         fprintf( outfile, "\tmovl %%ebp, %d(%%edi)\n", CONTEXTOFFSET(Ebp) );
 523         fprintf( outfile, "\tmovl %%esi, %d(%%edi)\n", CONTEXTOFFSET(Esp) );
 524                  /* The return glue code saved %esp into %esi */
 525     }
 526
 527     /* Restore the 32-bit registers */
 528     fprintf( outfile, "\tpopl %%edi\n" );
 529     fprintf( outfile, "\tpopl %%esi\n" );
 530     fprintf( outfile, "\tpopl %%ebx\n" );
 531
 532     /* Function exit sequence */
 533     fprintf( outfile, "\tpopl %%ebp\n" );
 534     fprintf( outfile, "\tret $12\n" );
 535
 536
 537     /* Start of the actual CallTo16 routine */
 538
 539     fprintf( outfile, ".L%s:\n", name );
 540
 541     /* Switch to the 16-bit stack */
 542     fprintf( outfile, "\tmovl %%esp,%%edx\n" );
 543     fprintf( outfile, "%s\t.byte 0x64\n\tmovw (%d),%%ss\n", data16_prefix(), STACKOFFSET + 2);
 544     fprintf( outfile, "\t.byte 0x64\n\tmovw (%d),%%sp\n", STACKOFFSET );
 545     fprintf( outfile, "\t.byte 0x64\n\tmovl %%edx,(%d)\n", STACKOFFSET );
 546
 547     /* Make %bp point to the previous stackframe (built by CallFrom16) */
 548     fprintf( outfile, "\tmovzwl %%sp,%%ebp\n" );
 549     fprintf( outfile, "\tleal %d(%%ebp),%%ebp\n", STACK16OFFSET(bp) );
 550
 551     /* Add the specified offset to the new sp */
 552     fprintf( outfile, "\tsubw %d(%%edx), %%sp\n", STACK32OFFSET(nb_args) );
 553
 554     if (reg_func)
 555     {
 556         /* Push the called routine address */
 557         fprintf( outfile, "\tmovl %d(%%edx),%%edx\n", STACK32OFFSET(target) );
 558         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(SegCs) );
 559         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(Eip) );
 560
 561         /* Get the registers */
 562         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(SegDs) );
 563         fprintf( outfile, "\tpushl %d(%%edx)\n", CONTEXTOFFSET(SegEs) );
 564         fprintf( outfile, "\tpopl %%es\n" );
 565         fprintf( outfile, "\tpushl %d(%%edx)\n", CONTEXTOFFSET(SegFs) );
 566         fprintf( outfile, "\tpopl %%fs\n" );
 567         fprintf( outfile, "\tpushl %d(%%edx)\n", CONTEXTOFFSET(SegGs) );
 568         fprintf( outfile, "\tpopl %%gs\n" );
 569         fprintf( outfile, "\tmovl %d(%%edx),%%ebp\n", CONTEXTOFFSET(Ebp) );
 570         fprintf( outfile, "\tmovl %d(%%edx),%%esi\n", CONTEXTOFFSET(Esi) );
 571         fprintf( outfile, "\tmovl %d(%%edx),%%edi\n", CONTEXTOFFSET(Edi) );
 572         fprintf( outfile, "\tmovl %d(%%edx),%%eax\n", CONTEXTOFFSET(Eax) );
 573         fprintf( outfile, "\tmovl %d(%%edx),%%ebx\n", CONTEXTOFFSET(Ebx) );
 574         fprintf( outfile, "\tmovl %d(%%edx),%%ecx\n", CONTEXTOFFSET(Ecx) );
 575         fprintf( outfile, "\tmovl %d(%%edx),%%edx\n", CONTEXTOFFSET(Edx) );
 576
 577         /* Get the 16-bit ds */
 578         fprintf( outfile, "\tpopw %%ds\n" );
 579     }
 580     else  /* not a register function */
 581     {
 582         /* Push the called routine address */
 583         fprintf( outfile, "\tpushl %d(%%edx)\n", STACK32OFFSET(target) );
 584
 585         /* Set %fs and %gs to the value saved by the last CallFrom16 */
 586         fprintf( outfile, "\tpushw %d(%%ebp)\n", STACK16OFFSET(fs)-STACK16OFFSET(bp) );
 587         fprintf( outfile, "\tpopw %%fs\n" );
 588         fprintf( outfile, "\tpushw %d(%%ebp)\n", STACK16OFFSET(gs)-STACK16OFFSET(bp) );
 589         fprintf( outfile, "\tpopw %%gs\n" );
 590
 591         /* Set %ds and %es (and %ax just in case) equal to %ss */
 592         fprintf( outfile, "\tmovw %%ss,%%ax\n" );
 593         fprintf( outfile, "\tmovw %%ax,%%ds\n" );
 594         fprintf( outfile, "\tmovw %%ax,%%es\n" );
 595     }
 596
 597     /* Jump to the called routine */
 598     fprintf( outfile, "\t.byte 0x66\n" );
 599     fprintf( outfile, "\tlret\n" );
 600
 601     /* Function footer */
 602     function_footer( outfile, name );
 603 }
 604
 605
 606 /*******************************************************************
 607  *         BuildRet16Func
 608  *
 609  * Build the return code for 16-bit callbacks
 610  */
 611 static void BuildRet16Func( FILE *outfile )
 612 {
 613     function_header( outfile, "CallTo16_Ret" );
 614
 615     /* Save %esp into %esi */
 616     fprintf( outfile, "\tmovl %%esp,%%esi\n" );
 617
 618     /* Restore 32-bit segment registers */
 619
 620     fprintf( outfile, "\t.byte 0x2e\n\tmovl %s", asm_name("CallTo16_DataSelector") );
 621     fprintf( outfile, "-%s,%%edi\n", asm_name("Call16_Ret_Start") );
 622     fprintf( outfile, "%s\tmovw %%di,%%ds\n", data16_prefix() );
 623     fprintf( outfile, "%s\tmovw %%di,%%es\n", data16_prefix() );
 624
 625     fprintf( outfile, "\t.byte 0x2e\n\tmov %s", asm_name("CallTo16_TebSelector") );
 626     fprintf( outfile, "-%s,%%fs\n", asm_name("Call16_Ret_Start") );
 627
 628     fprintf( outfile, "\t.byte 0x64\n\tmov (%d),%%gs\n", STRUCTOFFSET(TEB,gs_sel) );
 629
 630     /* Restore the 32-bit stack */
 631
 632     fprintf( outfile, "%s\tmovw %%di,%%ss\n", data16_prefix() );
 633     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d),%%esp\n", STACKOFFSET );
 634
 635     /* Return to caller */
 636
 637     fprintf( outfile, "\tlret\n" );
 638
 639     /* Function footer */
 640     function_footer( outfile, "CallTo16_Ret" );
 641
 642     /* Declare the return address and data selector variables */
 643
 644     fprintf( outfile, "\n\t.align %d\n", get_alignment(4) );
 645     fprintf( outfile, "\t.globl %s\n", asm_name("CallTo16_DataSelector") );
 646     fprintf( outfile, "%s:\t.long 0\n", asm_name("CallTo16_DataSelector") );
 647     fprintf( outfile, "\t.globl %s\n", asm_name("CallTo16_TebSelector") );
 648     fprintf( outfile, "%s:\t.long 0\n", asm_name("CallTo16_TebSelector") );
 649 }
 650
 651
 652 /*******************************************************************
 653  *         BuildCallTo32CBClient
 654  *
 655  * Call a CBClient relay stub from 32-bit code (KERNEL.620).
 656  *
 657  * Since the relay stub is itself 32-bit, this should not be a problem;
 658  * unfortunately, the relay stubs are expected to switch back to a
 659  * 16-bit stack (and 16-bit code) after completion :-(
 660  *
 661  * This would conflict with our 16- vs. 32-bit stack handling, so
 662  * we simply switch *back* to our 32-bit stack before returning to
 663  * the caller ...
 664  *
 665  * The CBClient relay stub expects to be called with the following
 666  * 16-bit stack layout, and with ebp and ebx pointing into the 16-bit
 667  * stack at the designated places:
 668  *
 669  *    ...
 670  *  (ebp+14) original arguments to the callback routine
 671  *  (ebp+10) far return address to original caller
 672  *  (ebp+6)  Thunklet target address
 673  *  (ebp+2)  Thunklet relay ID code
 674  *  (ebp)    BP (saved by CBClientGlueSL)
 675  *  (ebp-2)  SI (saved by CBClientGlueSL)
 676  *  (ebp-4)  DI (saved by CBClientGlueSL)
 677  *  (ebp-6)  DS (saved by CBClientGlueSL)
 678  *
 679  *   ...     buffer space used by the 16-bit side glue for temp copies
 680  *
 681  *  (ebx+4)  far return address to 16-bit side glue code
 682  *  (ebx)    saved 16-bit ss:sp (pointing to ebx+4)
 683  *
 684  * The 32-bit side glue code accesses both the original arguments (via ebp)
 685  * and the temporary copies prepared by the 16-bit side glue (via ebx).
 686  * After completion, the stub will load ss:sp from the buffer at ebx
 687  * and perform a far return to 16-bit code.
 688  *
 689  * To trick the relay stub into returning to us, we replace the 16-bit
 690  * return address to the glue code by a cs:ip pair pointing to our
 691  * return entry point (the original return address is saved first).
 692  * Our return stub thus called will then reload the 32-bit ss:esp and
 693  * return to 32-bit code (by using and ss:esp value that we have also
 694  * pushed onto the 16-bit stack before and a cs:eip values found at
 695  * that position on the 32-bit stack).  The ss:esp to be restored is
 696  * found relative to the 16-bit stack pointer at:
 697  *
 698  *  (ebx-4)   ss  (flat)
 699  *  (ebx-8)   sp  (32-bit stack pointer)
 700  *
 701  * The second variant of this routine, CALL32_CBClientEx, which is used
 702  * to implement KERNEL.621, has to cope with yet another problem: Here,
 703  * the 32-bit side directly returns to the caller of the CBClient thunklet,
 704  * restoring registers saved by CBClientGlueSL and cleaning up the stack.
 705  * As we have to return to our 32-bit code first, we have to adapt the
 706  * layout of our temporary area so as to include values for the registers
 707  * that are to be restored, and later (in the implementation of KERNEL.621)
 708  * we *really* restore them. The return stub restores DS, DI, SI, and BP
 709  * from the stack, skips the next 8 bytes (CBClient relay code / target),
 710  * and then performs a lret NN, where NN is the number of arguments to be
 711  * removed. Thus, we prepare our temporary area as follows:
 712  *
 713  *     (ebx+22) 16-bit cs  (this segment)
 714  *     (ebx+20) 16-bit ip  ('16-bit' return entry point)
 715  *     (ebx+16) 32-bit ss  (flat)
 716  *     (ebx+12) 32-bit sp  (32-bit stack pointer)
 717  *     (ebx+10) 16-bit bp  (points to ebx+24)
 718  *     (ebx+8)  16-bit si  (ignored)
 719  *     (ebx+6)  16-bit di  (ignored)
 720  *     (ebx+4)  16-bit ds  (we actually use the flat DS here)
 721  *     (ebx+2)  16-bit ss  (16-bit stack segment)
 722  *     (ebx+0)  16-bit sp  (points to ebx+4)
 723  *
 724  * Note that we ensure that DS is not changed and remains the flat segment,
 725  * and the 32-bit stack pointer our own return stub needs fits just
 726  * perfectly into the 8 bytes that are skipped by the Windows stub.
 727  * One problem is that we have to determine the number of removed arguments,
 728  * as these have to be really removed in KERNEL.621. Thus, the BP value
 729  * that we place in the temporary area to be restored, contains the value
 730  * that SP would have if no arguments were removed. By comparing the actual
 731  * value of SP with this value in our return stub we can compute the number
 732  * of removed arguments. This is then returned to KERNEL.621.
 733  *
 734  * The stack layout of this function:
 735  * (ebp+20)  nArgs     pointer to variable receiving nr. of args (Ex only)
 736  * (ebp+16)  esi       pointer to caller's esi value
 737  * (ebp+12)  arg       ebp value to be set for relay stub
 738  * (ebp+8)   func      CBClient relay stub address
 739  * (ebp+4)   ret addr
 740  * (ebp)     ebp
 741  */
 742 static void BuildCallTo32CBClient( FILE *outfile, BOOL isEx )
 743 {
 744     const char *name = isEx? "CALL32_CBClientEx" : "CALL32_CBClient";
 745     int size = isEx? 24 : 12;
 746
 747     /* Function header */
 748
 749     fprintf( outfile, "\n\t.align %d\n", get_alignment(4) );
 750     fprintf( outfile, "\t.globl %s\n", asm_name(name) );
 751     fprintf( outfile, "%s:\n", asm_name(name) );
 752
 753     /* Entry code */
 754
 755     fprintf( outfile, "\tpushl %%ebp\n" );
 756     fprintf( outfile, "\tmovl %%esp,%%ebp\n" );
 757     fprintf( outfile, "\tpushl %%edi\n" );
 758     fprintf( outfile, "\tpushl %%esi\n" );
 759     fprintf( outfile, "\tpushl %%ebx\n" );
 760
 761     if (UsePIC)
 762     {
 763         /* Get Global Offset Table into %edx */
 764         fprintf( outfile, "\tcall .L__wine_%s.getgot1\n", name );
 765         fprintf( outfile, ".L__wine_%s.getgot1:\n", name );
 766         fprintf( outfile, "\tpopl %%edx\n" );
 767         fprintf( outfile, "\taddl $_GLOBAL_OFFSET_TABLE_+[.-.L__wine_%s.getgot1], %%edx\n", name );
 768     }
 769
 770     /* Get the 16-bit stack */
 771
 772     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d),%%ebx\n", STACKOFFSET);
 773
 774     /* Convert it to a flat address */
 775
 776     fprintf( outfile, "\tshldl $16,%%ebx,%%eax\n" );
 777     fprintf( outfile, "\tandl $0xfff8,%%eax\n" );
 778     fprintf( outfile, "\tshrl $1,%%eax\n" );
 779     if (!UsePIC)
 780         fprintf( outfile, "\tmovl %s(%%eax),%%esi\n", asm_name("wine_ldt_copy") );
 781     else
 782     {
 783         fprintf( outfile, "\tmovl %s(%%edx), %%esi\n", asm_name("wine_ldt_copy@GOT") );
 784         fprintf( outfile, "\tmovl (%%esi,%%eax), %%esi\n" );
 785     }
 786     fprintf( outfile, "\tmovw %%bx,%%ax\n" );
 787     fprintf( outfile, "\taddl %%eax,%%esi\n" );
 788
 789     /* Allocate temporary area (simulate STACK16_PUSH) */
 790
 791     fprintf( outfile, "\tpushf\n" );
 792     fprintf( outfile, "\tcld\n" );
 793     fprintf( outfile, "\tleal -%d(%%esi), %%edi\n", size );
 794     fprintf( outfile, "\tmovl $%d, %%ecx\n", sizeof(STACK16FRAME) );
 795     fprintf( outfile, "\trep\n\tmovsb\n" );
 796     fprintf( outfile, "\tpopf\n" );
 797
 798     fprintf( outfile, "\t.byte 0x64\n\tsubw $%d,(%d)\n", size, STACKOFFSET );
 799
 800     fprintf( outfile, "\tpushl %%edi\n" );  /* remember address */
 801
 802     /* Set up temporary area */
 803
 804     if ( !isEx )
 805     {
 806         fprintf( outfile, "\tleal 4(%%edi), %%edi\n" );
 807
 808         fprintf( outfile, "\tleal -8(%%esp), %%eax\n" );
 809         fprintf( outfile, "\tmovl %%eax, -8(%%edi)\n" );    /* 32-bit sp */
 810
 811         fprintf( outfile, "\tmovw %%ss, %%ax\n" );
 812         fprintf( outfile, "\tandl $0x0000ffff, %%eax\n" );
 813         fprintf( outfile, "\tmovl %%eax, -4(%%edi)\n" );    /* 32-bit ss */
 814
 815         fprintf( outfile, "\taddl $%d, %%ebx\n", sizeof(STACK16FRAME)-size+4 + 4 );
 816         fprintf( outfile, "\tmovl %%ebx, 0(%%edi)\n" );    /* 16-bit ss:sp */
 817
 818         if (!UsePIC)
 819             fprintf( outfile, "\tmovl %s_RetAddr, %%eax\n", asm_name(name) );
 820         else
 821         {
 822             fprintf( outfile, "\tmovl %s_RetAddr@GOT(%%edx), %%eax\n", asm_name(name) );
 823             fprintf( outfile, "\tmovl (%%eax), %%eax\n" );
 824         }
 825         fprintf( outfile, "\tmovl %%eax, 4(%%edi)\n" );   /* overwrite return address */
 826     }
 827     else
 828     {
 829         fprintf( outfile, "\taddl $%d, %%ebx\n", sizeof(STACK16FRAME)-size+4 );
 830         fprintf( outfile, "\tmovl %%ebx, 0(%%edi)\n" );
 831
 832         fprintf( outfile, "\tmovw %%ds, %%ax\n" );
 833         fprintf( outfile, "\tmovw %%ax, 4(%%edi)\n" );
 834
 835         fprintf( outfile, "\taddl $20, %%ebx\n" );
 836         fprintf( outfile, "\tmovw %%bx, 10(%%edi)\n" );
 837
 838         fprintf( outfile, "\tleal -8(%%esp), %%eax\n" );
 839         fprintf( outfile, "\tmovl %%eax, 12(%%edi)\n" );
 840
 841         fprintf( outfile, "\tmovw %%ss, %%ax\n" );
 842         fprintf( outfile, "\tandl $0x0000ffff, %%eax\n" );
 843         fprintf( outfile, "\tmovl %%eax, 16(%%edi)\n" );
 844
 845         if (!UsePIC)
 846             fprintf( outfile, "\tmovl %s_RetAddr, %%eax\n", asm_name(name) );
 847         else
 848         {
 849             fprintf( outfile, "\tmovl %s_RetAddr@GOT(%%edx), %%eax\n", asm_name(name) );
 850             fprintf( outfile, "\tmovl (%%eax), %%eax\n" );
 851         }
 852         fprintf( outfile, "\tmovl %%eax, 20(%%edi)\n" );
 853     }
 854
 855     /* Set up registers and call CBClient relay stub (simulating a far call) */
 856
 857     fprintf( outfile, "\tmovl 16(%%ebp), %%esi\n" );
 858     fprintf( outfile, "\tmovl (%%esi), %%esi\n" );
 859
 860     fprintf( outfile, "\tmovl %%edi, %%ebx\n" );
 861     fprintf( outfile, "\tmovl 8(%%ebp), %%eax\n" );
 862     fprintf( outfile, "\tmovl 12(%%ebp), %%ebp\n" );
 863
 864     fprintf( outfile, "\tpushl %%cs\n" );
 865     fprintf( outfile, "\tcall *%%eax\n" );
 866
 867     /* Return new esi value to caller */
 868
 869     fprintf( outfile, "\tmovl 32(%%esp), %%edi\n" );
 870     fprintf( outfile, "\tmovl %%esi, (%%edi)\n" );
 871
 872     /* Cleanup temporary area (simulate STACK16_POP) */
 873
 874     fprintf( outfile, "\tpop %%esi\n" );
 875
 876     fprintf( outfile, "\tpushf\n" );
 877     fprintf( outfile, "\tstd\n" );
 878     fprintf( outfile, "\tdec %%esi\n" );
 879     fprintf( outfile, "\tleal %d(%%esi), %%edi\n", size );
 880     fprintf( outfile, "\tmovl $%d, %%ecx\n", sizeof(STACK16FRAME) );
 881     fprintf( outfile, "\trep\n\tmovsb\n" );
 882     fprintf( outfile, "\tpopf\n" );
 883
 884     fprintf( outfile, "\t.byte 0x64\n\taddw $%d,(%d)\n", size, STACKOFFSET );
 885
 886     /* Return argument size to caller */
 887     if ( isEx )
 888     {
 889         fprintf( outfile, "\tmovl 32(%%esp), %%ebx\n" );
 890         fprintf( outfile, "\tmovl %%ebp, (%%ebx)\n" );
 891     }
 892
 893     /* Restore registers and return */
 894
 895     fprintf( outfile, "\tpopl %%ebx\n" );
 896     fprintf( outfile, "\tpopl %%esi\n" );
 897     fprintf( outfile, "\tpopl %%edi\n" );
 898     fprintf( outfile, "\tpopl %%ebp\n" );
 899     fprintf( outfile, "\tret\n" );
 900     function_footer( outfile, name );
 901 }
 902
 903 static void BuildCallTo32CBClientRet( FILE *outfile, BOOL isEx )
 904 {
 905     const char *name = isEx? "CALL32_CBClientEx_Ret" : "CALL32_CBClient_Ret";
 906
 907     /* '16-bit' return stub */
 908
 909     fprintf( outfile, "\n\t.globl %s\n", asm_name(name) );
 910     fprintf( outfile, "%s:\n", asm_name(name) );
 911
 912     if ( !isEx )
 913     {
 914         fprintf( outfile, "\tmovzwl %%sp, %%ebx\n" );
 915         fprintf( outfile, "\tlssl %%ss:-16(%%ebx), %%esp\n" );
 916     }
 917     else
 918     {
 919         fprintf( outfile, "\tmovzwl %%bp, %%ebx\n" );
 920         fprintf( outfile, "\tsubw %%bp, %%sp\n" );
 921         fprintf( outfile, "\tmovzwl %%sp, %%ebp\n" );
 922         fprintf( outfile, "\tlssl %%ss:-12(%%ebx), %%esp\n" );
 923     }
 924     fprintf( outfile, "\tlret\n" );
 925     function_footer( outfile, name );
 926
 927     /* Declare the return address variable */
 928
 929     fprintf( outfile, "\n\t.globl %sAddr\n", asm_name(name) );
 930     fprintf( outfile, "%sAddr:\t.long 0\n", asm_name(name) );
 931 }
 932
 933
 934 /*******************************************************************
 935  *         BuildCallFrom32Regs
 936  *
 937  * Build a 32-bit-to-Wine call-back function for a 'register' function.
 938  * 'args' is the number of dword arguments.
 939  *
 940  * Stack layout:
 941  *   ...
 942  * (ebp+12)  first arg
 943  * (ebp+8)   ret addr to user code
 944  * (ebp+4)   ret addr to relay code
 945  * (ebp+0)   saved ebp
 946  * (ebp-128) buffer area to allow stack frame manipulation
 947  * (ebp-332) CONTEXT86 struct
 948  * (ebp-336) CONTEXT86 *argument
 949  *  ....     other arguments copied from (ebp+12)
 950  *
 951  * The entry point routine is called with a CONTEXT* extra argument,
 952  * following the normal args. In this context structure, EIP_reg
 953  * contains the return address to user code, and ESP_reg the stack
 954  * pointer on return (with the return address and arguments already
 955  * removed).
 956  */
 957 static void BuildCallFrom32Regs( FILE *outfile )
 958 {
 959     static const int STACK_SPACE = 128 + sizeof(CONTEXT86);
 960
 961     /* Function header */
 962
 963     function_header( outfile, "__wine_call_from_32_regs" );
 964
 965     /* Allocate some buffer space on the stack */
 966
 967     fprintf( outfile, "\tpushl %%ebp\n" );
 968     fprintf( outfile, "\tmovl %%esp,%%ebp\n ");
 969     fprintf( outfile, "\tleal -%d(%%esp), %%esp\n", STACK_SPACE );
 970
 971     /* Build the context structure */
 972
 973     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Eax) - STACK_SPACE );
 974     fprintf( outfile, "\tpushfl\n" );
 975     fprintf( outfile, "\tpopl %%eax\n" );
 976     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(EFlags) - STACK_SPACE );
 977     fprintf( outfile, "\tmovl 0(%%ebp),%%eax\n" );
 978     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Ebp) - STACK_SPACE );
 979     fprintf( outfile, "\tmovl %%ebx,%d(%%ebp)\n", CONTEXTOFFSET(Ebx) - STACK_SPACE );
 980     fprintf( outfile, "\tmovl %%ecx,%d(%%ebp)\n", CONTEXTOFFSET(Ecx) - STACK_SPACE );
 981     fprintf( outfile, "\tmovl %%edx,%d(%%ebp)\n", CONTEXTOFFSET(Edx) - STACK_SPACE );
 982     fprintf( outfile, "\tmovl %%esi,%d(%%ebp)\n", CONTEXTOFFSET(Esi) - STACK_SPACE );
 983     fprintf( outfile, "\tmovl %%edi,%d(%%ebp)\n", CONTEXTOFFSET(Edi) - STACK_SPACE );
 984
 985     fprintf( outfile, "\txorl %%eax,%%eax\n" );
 986     fprintf( outfile, "\tmovw %%cs,%%ax\n" );
 987     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegCs) - STACK_SPACE );
 988     fprintf( outfile, "\tmovw %%es,%%ax\n" );
 989     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegEs) - STACK_SPACE );
 990     fprintf( outfile, "\tmovw %%fs,%%ax\n" );
 991     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegFs) - STACK_SPACE );
 992     fprintf( outfile, "\tmovw %%gs,%%ax\n" );
 993     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegGs) - STACK_SPACE );
 994     fprintf( outfile, "\tmovw %%ss,%%ax\n" );
 995     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegSs) - STACK_SPACE );
 996     fprintf( outfile, "\tmovw %%ds,%%ax\n" );
 997     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegDs) - STACK_SPACE );
 998     fprintf( outfile, "\tmovw %%ax,%%es\n" );  /* set %es equal to %ds just in case */
 999
1000     fprintf( outfile, "\tmovl $0x%x,%%eax\n", CONTEXT86_FULL );
1001     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(ContextFlags) - STACK_SPACE );
1002
1003     fprintf( outfile, "\tmovl 8(%%ebp),%%eax\n" ); /* Get %eip at time of call */
1004     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Eip) - STACK_SPACE );
1005
1006     /* Transfer the arguments */
1007
1008     fprintf( outfile, "\tmovl 4(%%ebp),%%ebx\n" );   /* get relay code addr */
1009     fprintf( outfile, "\tpushl %%esp\n" );           /* push ptr to context struct */
1010     fprintf( outfile, "\tmovzbl 4(%%ebx),%%ecx\n" ); /* fetch number of args to copy */
1011     fprintf( outfile, "\tjecxz 1f\n" );
1012     fprintf( outfile, "\tsubl %%ecx,%%esp\n" );
1013     fprintf( outfile, "\tleal 12(%%ebp),%%esi\n" );  /* get %esp at time of call */
1014     fprintf( outfile, "\tmovl %%esp,%%edi\n" );
1015     fprintf( outfile, "\tshrl $2,%%ecx\n" );
1016     fprintf( outfile, "\tcld\n" );
1017     fprintf( outfile, "\trep\n\tmovsl\n" );  /* copy args */
1018
1019     fprintf( outfile, "1:\tmovzbl 5(%%ebx),%%eax\n" ); /* fetch number of args to remove */
1020     fprintf( outfile, "\tleal 12(%%ebp,%%eax),%%eax\n" );
1021     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Esp) - STACK_SPACE );
1022
1023     /* Call the entry point */
1024
1025     fprintf( outfile, "\taddl (%%ebx),%%ebx\n" );
1026     fprintf( outfile, "\tcall *%%ebx\n" );
1027     fprintf( outfile, "\tleal -%d(%%ebp),%%ecx\n", STACK_SPACE );
1028
1029     /* Restore the context structure */
1030
1031     fprintf( outfile, "2:\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegEs) );
1032     fprintf( outfile, "\tpopl %%es\n" );
1033     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegFs) );
1034     fprintf( outfile, "\tpopl %%fs\n" );
1035     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegGs) );
1036     fprintf( outfile, "\tpopl %%gs\n" );
1037
1038     fprintf( outfile, "\tmovl %d(%%ecx),%%edi\n", CONTEXTOFFSET(Edi) );
1039     fprintf( outfile, "\tmovl %d(%%ecx),%%esi\n", CONTEXTOFFSET(Esi) );
1040     fprintf( outfile, "\tmovl %d(%%ecx),%%edx\n", CONTEXTOFFSET(Edx) );
1041     fprintf( outfile, "\tmovl %d(%%ecx),%%ebx\n", CONTEXTOFFSET(Ebx) );
1042     fprintf( outfile, "\tmovl %d(%%ecx),%%eax\n", CONTEXTOFFSET(Eax) );
1043     fprintf( outfile, "\tmovl %d(%%ecx),%%ebp\n", CONTEXTOFFSET(Ebp) );
1044
1045     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegSs) );
1046     fprintf( outfile, "\tpopl %%ss\n" );
1047     fprintf( outfile, "\tmovl %d(%%ecx),%%esp\n", CONTEXTOFFSET(Esp) );
1048
1049     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(EFlags) );
1050     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegCs) );
1051     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(Eip) );
1052     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegDs) );
1053     fprintf( outfile, "\tmovl %d(%%ecx),%%ecx\n", CONTEXTOFFSET(Ecx) );
1054
1055     fprintf( outfile, "\tpopl %%ds\n" );
1056     fprintf( outfile, "\tiret\n" );
1057     function_footer( outfile, "__wine_call_from_32_regs" );
1058
1059     function_header( outfile, "__wine_call_from_32_restore_regs" );
1060     fprintf( outfile, "\tleal 4(%%esp),%%ecx\n" );
1061     fprintf( outfile, "\tjmp 2b\n" );
1062     function_footer( outfile, "__wine_call_from_32_restore_regs" );
1063 }
1064
1065
1066 /*******************************************************************
1067  *         BuildPendingEventCheck
1068  *
1069  * Build a function that checks whether there are any
1070  * pending DPMI events.
1071  *
1072  * Stack layout:
1073  *
1074  * (sp+12) long   eflags
1075  * (sp+6)  long   cs
1076  * (sp+2)  long   ip
1077  * (sp)    word   fs
1078  *
1079  * On entry to function, fs register points to a valid TEB.
1080  * On exit from function, stack will be popped.
1081  */
1082 static void BuildPendingEventCheck( FILE *outfile )
1083 {
1084     /* Function header */
1085
1086     function_header( outfile, "DPMI_PendingEventCheck" );
1087
1088     /* Check for pending events. */
1089
1090     fprintf( outfile, "\t.byte 0x64\n\ttestl $0xffffffff,(%d)\n",
1091              STRUCTOFFSET(TEB,vm86_pending) );
1092     fprintf( outfile, "\tje %s\n", asm_name("DPMI_PendingEventCheck_Cleanup") );
1093
1094     fprintf( outfile, "\t.byte 0x64\n\ttestl $0xffffffff,(%d)\n",
1095              STRUCTOFFSET(TEB,dpmi_vif) );
1096
1097     fprintf( outfile, "\tje %s\n", asm_name("DPMI_PendingEventCheck_Cleanup") );
1098
1099     /* Process pending events. */
1100
1101     fprintf( outfile, "\tsti\n" );
1102
1103     /* Start cleanup. Restore fs register. */
1104
1105     fprintf( outfile, ".globl %s\n", asm_name("DPMI_PendingEventCheck_Cleanup") );
1106     fprintf( outfile, "%s:\n", asm_name("DPMI_PendingEventCheck_Cleanup") );
1107     fprintf( outfile, "\tpopw %%fs\n" );
1108
1109     /* Return from function. */
1110
1111     fprintf( outfile, ".globl %s\n", asm_name("DPMI_PendingEventCheck_Return") );
1112     fprintf( outfile, "%s:\n", asm_name("DPMI_PendingEventCheck_Return") );
1113     fprintf( outfile, "\tiret\n" );
1114
1115     function_footer( outfile, "DPMI_PendingEventCheck" );
1116 }
1117
1118
1119 /*******************************************************************
1120  *         BuildRelays16
1121  *
1122  * Build all the 16-bit relay callbacks
1123  */
1124 void BuildRelays16( FILE *outfile )
1125 {
1126     if (target_cpu != CPU_x86)
1127     {
1128         fprintf( outfile, "/* File not used with this architecture. Do not edit! */\n\n" );
1129         return;
1130     }
1131
1132     /* File header */
1133
1134     fprintf( outfile, "/* File generated automatically. Do not edit! */\n\n" );
1135     fprintf( outfile, "\t.text\n" );
1136
1137     fprintf( outfile, "%s:\n\n", asm_name("__wine_spec_thunk_text_16") );
1138
1139     fprintf( outfile, "\t.globl %s\n", asm_name("Call16_Start") );
1140     fprintf( outfile, "%s:\n", asm_name("Call16_Start") );
1141     fprintf( outfile, "\t.byte 0\n\n" );
1142
1143     /* Standard CallFrom16 routine (WORD return) */
1144     BuildCallFrom16Core( outfile, FALSE, FALSE, TRUE );
1145
1146     /* Standard CallFrom16 routine (DWORD return) */
1147     BuildCallFrom16Core( outfile, FALSE, FALSE, FALSE );
1148
1149     /* Register CallFrom16 routine */
1150     BuildCallFrom16Core( outfile, TRUE, FALSE, FALSE );
1151
1152     /* C16ThkSL CallFrom16 routine */
1153     BuildCallFrom16Core( outfile, FALSE, TRUE, FALSE );
1154
1155     /* Standard CallTo16 routine */
1156     BuildCallTo16Core( outfile, 0 );
1157
1158     /* Register CallTo16 routine */
1159     BuildCallTo16Core( outfile, 1 );
1160
1161     /* CBClientThunkSL routine */
1162     BuildCallTo32CBClient( outfile, FALSE );
1163
1164     /* CBClientThunkSLEx routine */
1165     BuildCallTo32CBClient( outfile, TRUE  );
1166
1167     fprintf( outfile, "\t.globl %s\n", asm_name("Call16_End") );
1168     fprintf( outfile, "%s:\n", asm_name("Call16_End") );
1169     function_footer( outfile, "__wine_spec_thunk_text_16" );
1170
1171     /* The whole Call16_Ret segment must lie within the .data section */
1172     fprintf( outfile, "\n\t.data\n" );
1173     fprintf( outfile, "%s:\n\n", asm_name("__wine_spec_thunk_data_16") );
1174     fprintf( outfile, "\t.globl %s\n", asm_name("Call16_Ret_Start") );
1175     fprintf( outfile, "%s:\n", asm_name("Call16_Ret_Start") );
1176
1177     /* Standard CallTo16 return stub */
1178     BuildRet16Func( outfile );
1179
1180     /* CBClientThunkSL return stub */
1181     BuildCallTo32CBClientRet( outfile, FALSE );
1182
1183     /* CBClientThunkSLEx return stub */
1184     BuildCallTo32CBClientRet( outfile, TRUE  );
1185
1186     /* Pending DPMI events check stub */
1187     BuildPendingEventCheck( outfile );
1188
1189     /* End of Call16_Ret segment */
1190     fprintf( outfile, "\n\t.globl %s\n", asm_name("Call16_Ret_End") );
1191     fprintf( outfile, "%s:\n", asm_name("Call16_Ret_End") );
1192     function_footer( outfile, "__wine_spec_thunk_data_16" );
1193 }
1194
1195 /*******************************************************************
1196  *         BuildRelays32
1197  *
1198  * Build all the 32-bit relay callbacks
1199  */
1200 void BuildRelays32( FILE *outfile )
1201 {
1202     if (target_cpu != CPU_x86)
1203     {
1204         fprintf( outfile, "/* File not used with this architecture. Do not edit! */\n\n" );
1205         return;
1206     }
1207
1208     /* File header */
1209
1210     fprintf( outfile, "/* File generated automatically. Do not edit! */\n\n" );
1211     fprintf( outfile, "\t.text\n" );
1212     fprintf( outfile, "%s:\n\n", asm_name("__wine_spec_thunk_text_32") );
1213
1214     /* 32-bit register entry point */
1215     BuildCallFrom32Regs( outfile );
1216
1217     function_footer( outfile, "__wine_spec_thunk_text_32" );
1218 }