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., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  23  */
  24
  25 #include "config.h"
  26 #include "wine/port.h"
  27
  28 #include <ctype.h>
  29
  30 #include "thread.h"
  31 #include "wine/winbase16.h"
  32
  33 #include "build.h"
  34
  35 /* offset of a structure field relative to the start of the struct */
  36 #define STRUCTOFFSET(type,field) ((int)FIELD_OFFSET(type,field))
  37
  38 /* offset of register relative to the start of the CONTEXT struct */
  39 #define CONTEXTOFFSET(reg)  STRUCTOFFSET(CONTEXT86,reg)
  40
  41 /* offset of register relative to the start of the STACK16FRAME struct */
  42 #define STACK16OFFSET(reg)  STRUCTOFFSET(STACK16FRAME,reg)
  43
  44 /* offset of register relative to the start of the STACK32FRAME struct */
  45 #define STACK32OFFSET(reg)  STRUCTOFFSET(STACK32FRAME,reg)
  46
  47 /* offset of the stack pointer relative to %fs:(0) */
  48 #define STACKOFFSET 0xc0  /* STRUCTOFFSET(TEB,WOW32Reserved) */
  49
  50 /* fix this if the ntdll_thread_regs structure is changed */
  51 #define GS_OFFSET  0x1b0  /* STRUCTOFFSET(TEB,SpareBytes1) + STRUCTOFFSET(ntdll_thread_regs,gs) */
  52
  53 static void function_header( FILE *outfile, const char *name )
  54 {
  55     fprintf( outfile, "\n\t.align %d\n", get_alignment(4) );
  56     fprintf( outfile, "\t%s\n", func_declaration(name) );
  57     fprintf( outfile, "%s\n", asm_globl(name) );
  58 }
  59
  60
  61 static inline const char *data16_prefix(void)
  62 {
  63     return (target_platform == PLATFORM_SVR4) ? "\tdata16\n" : "";
  64 }
  65
  66 /*******************************************************************
  67  *         BuildCallFrom16Core
  68  *
  69  * This routine builds the core routines used in 16->32 thunks:
  70  * CallFrom16Word, CallFrom16Long, CallFrom16Register, and CallFrom16Thunk.
  71  *
  72  * These routines are intended to be called via a far call (with 32-bit
  73  * operand size) from 16-bit code.  The 16-bit code stub must push %bp,
  74  * the 32-bit entry point to be called, and the argument conversion
  75  * routine to be used (see stack layout below).
  76  *
  77  * The core routine completes the STACK16FRAME on the 16-bit stack and
  78  * switches to the 32-bit stack.  Then, the argument conversion routine
  79  * is called; it gets passed the 32-bit entry point and a pointer to the
  80  * 16-bit arguments (on the 16-bit stack) as parameters. (You can either
  81  * use conversion routines automatically generated by BuildCallFrom16,
  82  * or write your own for special purposes.)
  83  *
  84  * The conversion routine must call the 32-bit entry point, passing it
  85  * the converted arguments, and return its return value to the core.
  86  * After the conversion routine has returned, the core switches back
  87  * to the 16-bit stack, converts the return value to the DX:AX format
  88  * (CallFrom16Long), and returns to the 16-bit call stub.  All parameters,
  89  * including %bp, are popped off the stack.
  90  *
  91  * The 16-bit call stub now returns to the caller, popping the 16-bit
  92  * arguments if necessary (pascal calling convention).
  93  *
  94  * In the case of a 'register' function, CallFrom16Register fills a
  95  * CONTEXT86 structure with the values all registers had at the point
  96  * the first instruction of the 16-bit call stub was about to be
  97  * executed.  A pointer to this CONTEXT86 is passed as third parameter
  98  * to the argument conversion routine, which typically passes it on
  99  * to the called 32-bit entry point.
 100  *
 101  * CallFrom16Thunk is a special variant used by the implementation of
 102  * the Win95 16->32 thunk functions C16ThkSL and C16ThkSL01 and is
 103  * implemented as follows:
 104  * On entry, the EBX register is set up to contain a flat pointer to the
 105  * 16-bit stack such that EBX+22 points to the first argument.
 106  * Then, the entry point is called, while EBP is set up to point
 107  * to the return address (on the 32-bit stack).
 108  * The called function returns with CX set to the number of bytes
 109  * to be popped of the caller's stack.
 110  *
 111  * Stack layout upon entry to the core routine (STACK16FRAME):
 112  *  ...           ...
 113  * (sp+24) word   first 16-bit arg
 114  * (sp+22) word   cs
 115  * (sp+20) word   ip
 116  * (sp+18) word   bp
 117  * (sp+14) long   32-bit entry point (reused for Win16 mutex recursion count)
 118  * (sp+12) word   ip of actual entry point (necessary for relay debugging)
 119  * (sp+8)  long   relay (argument conversion) function entry point
 120  * (sp+4)  long   cs of 16-bit entry point
 121  * (sp)    long   ip of 16-bit entry point
 122  *
 123  * Added on the stack:
 124  * (sp-2)  word   saved gs
 125  * (sp-4)  word   saved fs
 126  * (sp-6)  word   saved es
 127  * (sp-8)  word   saved ds
 128  * (sp-12) long   saved ebp
 129  * (sp-16) long   saved ecx
 130  * (sp-20) long   saved edx
 131  * (sp-24) long   saved previous stack
 132  */
 133 static void BuildCallFrom16Core( FILE *outfile, int reg_func, int thunk )
 134 {
 135     /* Function header */
 136     if (thunk) function_header( outfile, "__wine_call_from_16_thunk" );
 137     else if (reg_func) function_header( outfile, "__wine_call_from_16_regs" );
 138     else function_header( outfile, "__wine_call_from_16" );
 139
 140     /* Create STACK16FRAME (except STACK32FRAME link) */
 141     fprintf( outfile, "\tpushw %%gs\n" );
 142     fprintf( outfile, "\tpushw %%fs\n" );
 143     fprintf( outfile, "\tpushw %%es\n" );
 144     fprintf( outfile, "\tpushw %%ds\n" );
 145     fprintf( outfile, "\tpushl %%ebp\n" );
 146     fprintf( outfile, "\tpushl %%ecx\n" );
 147     fprintf( outfile, "\tpushl %%edx\n" );
 148
 149     /* Save original EFlags register */
 150     if (reg_func) fprintf( outfile, "\tpushfl\n" );
 151
 152     if ( UsePIC )
 153     {
 154         fprintf( outfile, "\tcall 1f\n" );
 155         fprintf( outfile, "1:\tpopl %%ecx\n" );
 156         fprintf( outfile, "\t.byte 0x2e\n\tmovl %s-1b(%%ecx),%%edx\n",
 157                  asm_name("CallTo16_DataSelector") );
 158     }
 159     else
 160         fprintf( outfile, "\t.byte 0x2e\n\tmovl %s,%%edx\n", asm_name("CallTo16_DataSelector") );
 161
 162     /* Load 32-bit segment registers */
 163     fprintf( outfile, "%s\tmovw %%dx, %%ds\n", data16_prefix() );
 164     fprintf( outfile, "%s\tmovw %%dx, %%es\n", data16_prefix() );
 165
 166     if ( UsePIC )
 167         fprintf( outfile, "\tmovw %s-1b(%%ecx), %%fs\n", asm_name("CallTo16_TebSelector") );
 168     else
 169         fprintf( outfile, "\tmovw %s, %%fs\n", asm_name("CallTo16_TebSelector") );
 170
 171     fprintf( outfile, "\t.byte 0x64\n\tmov (%d),%%gs\n", GS_OFFSET );
 172
 173     /* Translate STACK16FRAME base to flat offset in %edx */
 174     fprintf( outfile, "\tmovw %%ss, %%dx\n" );
 175     fprintf( outfile, "\tandl $0xfff8, %%edx\n" );
 176     fprintf( outfile, "\tshrl $1, %%edx\n" );
 177     if (UsePIC)
 178     {
 179         fprintf( outfile, "\taddl wine_ldt_copy_ptr-1b(%%ecx),%%edx\n" );
 180         fprintf( outfile, "\tmovl (%%edx), %%edx\n" );
 181     }
 182     else
 183         fprintf( outfile, "\tmovl %s(%%edx), %%edx\n", asm_name("wine_ldt_copy") );
 184     fprintf( outfile, "\tmovzwl %%sp, %%ebp\n" );
 185     fprintf( outfile, "\tleal %d(%%ebp,%%edx), %%edx\n", reg_func ? 0 : -4 );
 186
 187     /* Get saved flags into %ecx */
 188     if (reg_func) fprintf( outfile, "\tpopl %%ecx\n" );
 189
 190     /* Get the 32-bit stack pointer from the TEB and complete STACK16FRAME */
 191     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d), %%ebp\n", STACKOFFSET );
 192     fprintf( outfile, "\tpushl %%ebp\n" );
 193
 194     /* Switch stacks */
 195     fprintf( outfile, "%s\t.byte 0x64\n\tmovw %%ss, (%d)\n", data16_prefix(), STACKOFFSET + 2 );
 196     fprintf( outfile, "\t.byte 0x64\n\tmovw %%sp, (%d)\n", STACKOFFSET );
 197     fprintf( outfile, "\tpushl %%ds\n" );
 198     fprintf( outfile, "\tpopl %%ss\n" );
 199     fprintf( outfile, "\tmovl %%ebp, %%esp\n" );
 200     fprintf( outfile, "\taddl $%d, %%ebp\n", STACK32OFFSET(ebp) );
 201
 202
 203     /* At this point:
 204        STACK16FRAME is completely set up
 205        DS, ES, SS: flat data segment
 206        FS: current TEB
 207        ESP: points to last STACK32FRAME
 208        EBP: points to ebp member of last STACK32FRAME
 209        EDX: points to current STACK16FRAME
 210        ECX: contains saved flags
 211        all other registers: unchanged */
 212
 213     /* Special case: C16ThkSL stub */
 214     if ( thunk )
 215     {
 216         /* Set up registers as expected and call thunk */
 217         fprintf( outfile, "\tleal %d(%%edx), %%ebx\n", sizeof(STACK16FRAME)-22 );
 218         fprintf( outfile, "\tleal -4(%%esp), %%ebp\n" );
 219
 220         fprintf( outfile, "\tcall *%d(%%edx)\n", STACK16OFFSET(entry_point) );
 221
 222         /* Switch stack back */
 223         fprintf( outfile, "\t.byte 0x64\n\tmovw (%d), %%ss\n", STACKOFFSET+2 );
 224         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%esp\n", STACKOFFSET );
 225         fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 226
 227         /* Restore registers and return directly to caller */
 228         fprintf( outfile, "\taddl $8, %%esp\n" );
 229         fprintf( outfile, "\tpopl %%ebp\n" );
 230         fprintf( outfile, "\tpopw %%ds\n" );
 231         fprintf( outfile, "\tpopw %%es\n" );
 232         fprintf( outfile, "\tpopw %%fs\n" );
 233         fprintf( outfile, "\tpopw %%gs\n" );
 234         fprintf( outfile, "\taddl $20, %%esp\n" );
 235
 236         fprintf( outfile, "\txorb %%ch, %%ch\n" );
 237         fprintf( outfile, "\tpopl %%ebx\n" );
 238         fprintf( outfile, "\taddw %%cx, %%sp\n" );
 239         fprintf( outfile, "\tpush %%ebx\n" );
 240
 241         fprintf( outfile, "\t.byte 0x66\n" );
 242         fprintf( outfile, "\tlret\n" );
 243
 244         return;
 245     }
 246
 247
 248     /* Build register CONTEXT */
 249     if ( reg_func )
 250     {
 251         fprintf( outfile, "\tsubl $%d, %%esp\n", sizeof(CONTEXT86) );
 252
 253         fprintf( outfile, "\tmovl %%ecx, %d(%%esp)\n", CONTEXTOFFSET(EFlags) );
 254
 255         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Eax) );
 256         fprintf( outfile, "\tmovl %%ebx, %d(%%esp)\n", CONTEXTOFFSET(Ebx) );
 257         fprintf( outfile, "\tmovl %%esi, %d(%%esp)\n", CONTEXTOFFSET(Esi) );
 258         fprintf( outfile, "\tmovl %%edi, %d(%%esp)\n", CONTEXTOFFSET(Edi) );
 259
 260         fprintf( outfile, "\tmovl %d(%%edx), %%eax\n", STACK16OFFSET(ebp) );
 261         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Ebp) );
 262         fprintf( outfile, "\tmovl %d(%%edx), %%eax\n", STACK16OFFSET(ecx) );
 263         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Ecx) );
 264         fprintf( outfile, "\tmovl %d(%%edx), %%eax\n", STACK16OFFSET(edx) );
 265         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Edx) );
 266
 267         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(ds) );
 268         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegDs) );
 269         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(es) );
 270         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegEs) );
 271         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(fs) );
 272         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegFs) );
 273         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(gs) );
 274         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegGs) );
 275
 276         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(cs) );
 277         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegCs) );
 278         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(ip) );
 279         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Eip) );
 280
 281         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%eax\n", STACKOFFSET+2 );
 282         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegSs) );
 283         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%eax\n", STACKOFFSET );
 284         fprintf( outfile, "\taddl $%d, %%eax\n", STACK16OFFSET(ip) );
 285         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Esp) );
 286 #if 0
 287         fprintf( outfile, "\tfsave %d(%%esp)\n", CONTEXTOFFSET(FloatSave) );
 288 #endif
 289
 290         /* Push address of CONTEXT86 structure -- popped by the relay routine */
 291         fprintf( outfile, "\tmovl %%esp,%%eax\n" );
 292         fprintf( outfile, "\tandl $~15,%%esp\n" );
 293         fprintf( outfile, "\tsubl $4,%%esp\n" );
 294         fprintf( outfile, "\tpushl %%eax\n" );
 295     }
 296     else
 297     {
 298         fprintf( outfile, "\tsubl $8,%%esp\n" );
 299         fprintf( outfile, "\tandl $~15,%%esp\n" );
 300         fprintf( outfile, "\taddl $8,%%esp\n" );
 301     }
 302
 303     /* Call relay routine (which will call the API entry point) */
 304     fprintf( outfile, "\tleal %d(%%edx), %%eax\n", sizeof(STACK16FRAME) );
 305     fprintf( outfile, "\tpushl %%eax\n" );
 306     fprintf( outfile, "\tpushl %d(%%edx)\n", STACK16OFFSET(entry_point) );
 307     fprintf( outfile, "\tcall *%d(%%edx)\n", STACK16OFFSET(relay) );
 308
 309     if ( reg_func )
 310     {
 311         fprintf( outfile, "\tleal -%d(%%ebp), %%ebx\n",
 312                  sizeof(CONTEXT) + STACK32OFFSET(ebp) );
 313
 314         /* Switch stack back */
 315         fprintf( outfile, "\t.byte 0x64\n\tmovw (%d), %%ss\n", STACKOFFSET+2 );
 316         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%esp\n", STACKOFFSET );
 317         fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 318
 319         /* Get return address to CallFrom16 stub */
 320         fprintf( outfile, "\taddw $%d, %%sp\n", STACK16OFFSET(callfrom_ip)-4 );
 321         fprintf( outfile, "\tpopl %%eax\n" );
 322         fprintf( outfile, "\tpopl %%edx\n" );
 323
 324         /* Restore all registers from CONTEXT */
 325         fprintf( outfile, "\tmovw %d(%%ebx), %%ss\n", CONTEXTOFFSET(SegSs) );
 326         fprintf( outfile, "\tmovl %d(%%ebx), %%esp\n", CONTEXTOFFSET(Esp) );
 327         fprintf( outfile, "\taddl $4, %%esp\n" );  /* room for final return address */
 328
 329         fprintf( outfile, "\tpushw %d(%%ebx)\n", CONTEXTOFFSET(SegCs) );
 330         fprintf( outfile, "\tpushw %d(%%ebx)\n", CONTEXTOFFSET(Eip) );
 331         fprintf( outfile, "\tpushl %%edx\n" );
 332         fprintf( outfile, "\tpushl %%eax\n" );
 333         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(EFlags) );
 334         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(SegDs) );
 335
 336         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(SegEs) );
 337         fprintf( outfile, "\tpopl %%es\n" );
 338         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(SegFs) );
 339         fprintf( outfile, "\tpopl %%fs\n" );
 340         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(SegGs) );
 341         fprintf( outfile, "\tpopl %%gs\n" );
 342
 343         fprintf( outfile, "\tmovl %d(%%ebx), %%ebp\n", CONTEXTOFFSET(Ebp) );
 344         fprintf( outfile, "\tmovl %d(%%ebx), %%esi\n", CONTEXTOFFSET(Esi) );
 345         fprintf( outfile, "\tmovl %d(%%ebx), %%edi\n", CONTEXTOFFSET(Edi) );
 346         fprintf( outfile, "\tmovl %d(%%ebx), %%eax\n", CONTEXTOFFSET(Eax) );
 347         fprintf( outfile, "\tmovl %d(%%ebx), %%edx\n", CONTEXTOFFSET(Edx) );
 348         fprintf( outfile, "\tmovl %d(%%ebx), %%ecx\n", CONTEXTOFFSET(Ecx) );
 349         fprintf( outfile, "\tmovl %d(%%ebx), %%ebx\n", CONTEXTOFFSET(Ebx) );
 350
 351         fprintf( outfile, "\tpopl %%ds\n" );
 352         fprintf( outfile, "\tpopfl\n" );
 353         fprintf( outfile, "\tlret\n" );
 354     }
 355     else
 356     {
 357         /* Switch stack back */
 358         fprintf( outfile, "\t.byte 0x64\n\tmovw (%d), %%ss\n", STACKOFFSET+2 );
 359         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%esp\n", STACKOFFSET );
 360         fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 361
 362         /* Restore registers */
 363         fprintf( outfile, "\tpopl %%edx\n" );
 364         fprintf( outfile, "\tpopl %%ecx\n" );
 365         fprintf( outfile, "\tpopl %%ebp\n" );
 366         fprintf( outfile, "\tpopw %%ds\n" );
 367         fprintf( outfile, "\tpopw %%es\n" );
 368         fprintf( outfile, "\tpopw %%fs\n" );
 369         fprintf( outfile, "\tpopw %%gs\n" );
 370
 371         /* Return to return stub which will return to caller */
 372         fprintf( outfile, "\tlret $12\n" );
 373     }
 374     if (thunk) output_function_size( outfile, "__wine_call_from_16_thunk" );
 375     else if (reg_func) output_function_size( outfile, "__wine_call_from_16_regs" );
 376     else output_function_size( outfile, "__wine_call_from_16" );
 377 }
 378
 379
 380 /*******************************************************************
 381  *         BuildCallTo16Core
 382  *
 383  * This routine builds the core routines used in 32->16 thunks:
 384  *
 385  * extern DWORD WINAPI wine_call_to_16( FARPROC16 target, DWORD cbArgs, PEXCEPTION_HANDLER handler );
 386  * extern void WINAPI wine_call_to_16_regs( CONTEXT86 *context, DWORD cbArgs, PEXCEPTION_HANDLER handler );
 387  *
 388  * These routines can be called directly from 32-bit code.
 389  *
 390  * All routines expect that the 16-bit stack contents (arguments) and the
 391  * return address (segptr to CallTo16_Ret) were already set up by the
 392  * caller; nb_args must contain the number of bytes to be conserved.  The
 393  * 16-bit SS:SP will be set accordinly.
 394  *
 395  * All other registers are either taken from the CONTEXT86 structure
 396  * or else set to default values.  The target routine address is either
 397  * given directly or taken from the CONTEXT86.
 398  */
 399 static void BuildCallTo16Core( FILE *outfile, int reg_func )
 400 {
 401     const char *name = reg_func ? "wine_call_to_16_regs" : "wine_call_to_16";
 402
 403     /* Function header */
 404     function_header( outfile, name );
 405
 406     /* Function entry sequence */
 407     fprintf( outfile, "\tpushl %%ebp\n" );
 408     fprintf( outfile, "\tmovl %%esp, %%ebp\n" );
 409
 410     /* Save the 32-bit registers */
 411     fprintf( outfile, "\tpushl %%ebx\n" );
 412     fprintf( outfile, "\tpushl %%esi\n" );
 413     fprintf( outfile, "\tpushl %%edi\n" );
 414     fprintf( outfile, "\t.byte 0x64\n\tmov %%gs,(%d)\n", GS_OFFSET );
 415
 416     /* Setup exception frame */
 417     fprintf( outfile, "\t.byte 0x64\n\tpushl (%d)\n", STACKOFFSET );
 418     fprintf( outfile, "\tpushl 16(%%ebp)\n" ); /* handler */
 419     fprintf( outfile, "\t.byte 0x64\n\tpushl (0)\n" );
 420     fprintf( outfile, "\t.byte 0x64\n\tmovl %%esp,(0)\n" );
 421
 422     /* Call the actual CallTo16 routine (simulate a lcall) */
 423     fprintf( outfile, "\tpushl %%cs\n" );
 424     fprintf( outfile, "\tcall .L%s\n", name );
 425
 426     /* Remove exception frame */
 427     fprintf( outfile, "\t.byte 0x64\n\tpopl (0)\n" );
 428     fprintf( outfile, "\taddl $4, %%esp\n" );
 429     fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 430
 431     if ( !reg_func )
 432     {
 433         /* Convert return value */
 434         fprintf( outfile, "\tandl $0xffff,%%eax\n" );
 435         fprintf( outfile, "\tshll $16,%%edx\n" );
 436         fprintf( outfile, "\torl %%edx,%%eax\n" );
 437     }
 438     else
 439     {
 440         /*
 441          * Modify CONTEXT86 structure to contain new values
 442          *
 443          * NOTE:  We restore only EAX, EBX, EDX, EDX, EBP, and ESP.
 444          *        The segment registers as well as ESI and EDI should
 445          *        not be modified by a well-behaved 16-bit routine in
 446          *        any case.  [If necessary, we could restore them as well,
 447          *        at the cost of a somewhat less efficient return path.]
 448          */
 449
 450         fprintf( outfile, "\tmovl %d(%%esp), %%edi\n", STACK32OFFSET(target) - STACK32OFFSET(edi));
 451                 /* everything above edi has been popped already */
 452
 453         fprintf( outfile, "\tmovl %%eax, %d(%%edi)\n", CONTEXTOFFSET(Eax) );
 454         fprintf( outfile, "\tmovl %%ebx, %d(%%edi)\n", CONTEXTOFFSET(Ebx) );
 455         fprintf( outfile, "\tmovl %%ecx, %d(%%edi)\n", CONTEXTOFFSET(Ecx) );
 456         fprintf( outfile, "\tmovl %%edx, %d(%%edi)\n", CONTEXTOFFSET(Edx) );
 457         fprintf( outfile, "\tmovl %%ebp, %d(%%edi)\n", CONTEXTOFFSET(Ebp) );
 458         fprintf( outfile, "\tmovl %%esi, %d(%%edi)\n", CONTEXTOFFSET(Esp) );
 459                  /* The return glue code saved %esp into %esi */
 460     }
 461
 462     /* Restore the 32-bit registers */
 463     fprintf( outfile, "\tpopl %%edi\n" );
 464     fprintf( outfile, "\tpopl %%esi\n" );
 465     fprintf( outfile, "\tpopl %%ebx\n" );
 466
 467     /* Function exit sequence */
 468     fprintf( outfile, "\tpopl %%ebp\n" );
 469     fprintf( outfile, "\tret $12\n" );
 470
 471
 472     /* Start of the actual CallTo16 routine */
 473
 474     fprintf( outfile, ".L%s:\n", name );
 475
 476     /* Switch to the 16-bit stack */
 477     fprintf( outfile, "\tmovl %%esp,%%edx\n" );
 478     fprintf( outfile, "%s\t.byte 0x64\n\tmovw (%d),%%ss\n", data16_prefix(), STACKOFFSET + 2);
 479     fprintf( outfile, "\t.byte 0x64\n\tmovw (%d),%%sp\n", STACKOFFSET );
 480     fprintf( outfile, "\t.byte 0x64\n\tmovl %%edx,(%d)\n", STACKOFFSET );
 481
 482     /* Make %bp point to the previous stackframe (built by CallFrom16) */
 483     fprintf( outfile, "\tmovzwl %%sp,%%ebp\n" );
 484     fprintf( outfile, "\tleal %d(%%ebp),%%ebp\n", STACK16OFFSET(bp) );
 485
 486     /* Add the specified offset to the new sp */
 487     fprintf( outfile, "\tsubw %d(%%edx), %%sp\n", STACK32OFFSET(nb_args) );
 488
 489     if (reg_func)
 490     {
 491         /* Push the called routine address */
 492         fprintf( outfile, "\tmovl %d(%%edx),%%edx\n", STACK32OFFSET(target) );
 493         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(SegCs) );
 494         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(Eip) );
 495
 496         /* Get the registers */
 497         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(SegDs) );
 498         fprintf( outfile, "\tpushl %d(%%edx)\n", CONTEXTOFFSET(SegEs) );
 499         fprintf( outfile, "\tpopl %%es\n" );
 500         fprintf( outfile, "\tpushl %d(%%edx)\n", CONTEXTOFFSET(SegFs) );
 501         fprintf( outfile, "\tpopl %%fs\n" );
 502         fprintf( outfile, "\tpushl %d(%%edx)\n", CONTEXTOFFSET(SegGs) );
 503         fprintf( outfile, "\tpopl %%gs\n" );
 504         fprintf( outfile, "\tmovl %d(%%edx),%%ebp\n", CONTEXTOFFSET(Ebp) );
 505         fprintf( outfile, "\tmovl %d(%%edx),%%esi\n", CONTEXTOFFSET(Esi) );
 506         fprintf( outfile, "\tmovl %d(%%edx),%%edi\n", CONTEXTOFFSET(Edi) );
 507         fprintf( outfile, "\tmovl %d(%%edx),%%eax\n", CONTEXTOFFSET(Eax) );
 508         fprintf( outfile, "\tmovl %d(%%edx),%%ebx\n", CONTEXTOFFSET(Ebx) );
 509         fprintf( outfile, "\tmovl %d(%%edx),%%ecx\n", CONTEXTOFFSET(Ecx) );
 510         fprintf( outfile, "\tmovl %d(%%edx),%%edx\n", CONTEXTOFFSET(Edx) );
 511
 512         /* Get the 16-bit ds */
 513         fprintf( outfile, "\tpopw %%ds\n" );
 514     }
 515     else  /* not a register function */
 516     {
 517         /* Push the called routine address */
 518         fprintf( outfile, "\tpushl %d(%%edx)\n", STACK32OFFSET(target) );
 519
 520         /* Set %fs and %gs to the value saved by the last CallFrom16 */
 521         fprintf( outfile, "\tpushw %d(%%ebp)\n", STACK16OFFSET(fs)-STACK16OFFSET(bp) );
 522         fprintf( outfile, "\tpopw %%fs\n" );
 523         fprintf( outfile, "\tpushw %d(%%ebp)\n", STACK16OFFSET(gs)-STACK16OFFSET(bp) );
 524         fprintf( outfile, "\tpopw %%gs\n" );
 525
 526         /* Set %ds and %es (and %ax just in case) equal to %ss */
 527         fprintf( outfile, "\tmovw %%ss,%%ax\n" );
 528         fprintf( outfile, "\tmovw %%ax,%%ds\n" );
 529         fprintf( outfile, "\tmovw %%ax,%%es\n" );
 530     }
 531
 532     /* Jump to the called routine */
 533     fprintf( outfile, "\t.byte 0x66\n" );
 534     fprintf( outfile, "\tlret\n" );
 535
 536     /* Function footer */
 537     output_function_size( outfile, name );
 538 }
 539
 540
 541 /*******************************************************************
 542  *         BuildRet16Func
 543  *
 544  * Build the return code for 16-bit callbacks
 545  */
 546 static void BuildRet16Func( FILE *outfile )
 547 {
 548     function_header( outfile, "__wine_call_to_16_ret" );
 549
 550     /* Save %esp into %esi */
 551     fprintf( outfile, "\tmovl %%esp,%%esi\n" );
 552
 553     /* Restore 32-bit segment registers */
 554
 555     fprintf( outfile, "\t.byte 0x2e\n\tmovl %s", asm_name("CallTo16_DataSelector") );
 556     fprintf( outfile, "-%s,%%edi\n", asm_name("__wine_call16_start") );
 557     fprintf( outfile, "%s\tmovw %%di,%%ds\n", data16_prefix() );
 558     fprintf( outfile, "%s\tmovw %%di,%%es\n", data16_prefix() );
 559
 560     fprintf( outfile, "\t.byte 0x2e\n\tmov %s", asm_name("CallTo16_TebSelector") );
 561     fprintf( outfile, "-%s,%%fs\n", asm_name("__wine_call16_start") );
 562
 563     fprintf( outfile, "\t.byte 0x64\n\tmov (%d),%%gs\n", GS_OFFSET );
 564
 565     /* Restore the 32-bit stack */
 566
 567     fprintf( outfile, "%s\tmovw %%di,%%ss\n", data16_prefix() );
 568     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d),%%esp\n", STACKOFFSET );
 569
 570     /* Return to caller */
 571
 572     fprintf( outfile, "\tlret\n" );
 573     output_function_size( outfile, "__wine_call_to_16_ret" );
 574 }
 575
 576
 577 /*******************************************************************
 578  *         BuildCallTo32CBClient
 579  *
 580  * Call a CBClient relay stub from 32-bit code (KERNEL.620).
 581  *
 582  * Since the relay stub is itself 32-bit, this should not be a problem;
 583  * unfortunately, the relay stubs are expected to switch back to a
 584  * 16-bit stack (and 16-bit code) after completion :-(
 585  *
 586  * This would conflict with our 16- vs. 32-bit stack handling, so
 587  * we simply switch *back* to our 32-bit stack before returning to
 588  * the caller ...
 589  *
 590  * The CBClient relay stub expects to be called with the following
 591  * 16-bit stack layout, and with ebp and ebx pointing into the 16-bit
 592  * stack at the designated places:
 593  *
 594  *    ...
 595  *  (ebp+14) original arguments to the callback routine
 596  *  (ebp+10) far return address to original caller
 597  *  (ebp+6)  Thunklet target address
 598  *  (ebp+2)  Thunklet relay ID code
 599  *  (ebp)    BP (saved by CBClientGlueSL)
 600  *  (ebp-2)  SI (saved by CBClientGlueSL)
 601  *  (ebp-4)  DI (saved by CBClientGlueSL)
 602  *  (ebp-6)  DS (saved by CBClientGlueSL)
 603  *
 604  *   ...     buffer space used by the 16-bit side glue for temp copies
 605  *
 606  *  (ebx+4)  far return address to 16-bit side glue code
 607  *  (ebx)    saved 16-bit ss:sp (pointing to ebx+4)
 608  *
 609  * The 32-bit side glue code accesses both the original arguments (via ebp)
 610  * and the temporary copies prepared by the 16-bit side glue (via ebx).
 611  * After completion, the stub will load ss:sp from the buffer at ebx
 612  * and perform a far return to 16-bit code.
 613  *
 614  * To trick the relay stub into returning to us, we replace the 16-bit
 615  * return address to the glue code by a cs:ip pair pointing to our
 616  * return entry point (the original return address is saved first).
 617  * Our return stub thus called will then reload the 32-bit ss:esp and
 618  * return to 32-bit code (by using and ss:esp value that we have also
 619  * pushed onto the 16-bit stack before and a cs:eip values found at
 620  * that position on the 32-bit stack).  The ss:esp to be restored is
 621  * found relative to the 16-bit stack pointer at:
 622  *
 623  *  (ebx-4)   ss  (flat)
 624  *  (ebx-8)   sp  (32-bit stack pointer)
 625  *
 626  * The second variant of this routine, CALL32_CBClientEx, which is used
 627  * to implement KERNEL.621, has to cope with yet another problem: Here,
 628  * the 32-bit side directly returns to the caller of the CBClient thunklet,
 629  * restoring registers saved by CBClientGlueSL and cleaning up the stack.
 630  * As we have to return to our 32-bit code first, we have to adapt the
 631  * layout of our temporary area so as to include values for the registers
 632  * that are to be restored, and later (in the implementation of KERNEL.621)
 633  * we *really* restore them. The return stub restores DS, DI, SI, and BP
 634  * from the stack, skips the next 8 bytes (CBClient relay code / target),
 635  * and then performs a lret NN, where NN is the number of arguments to be
 636  * removed. Thus, we prepare our temporary area as follows:
 637  *
 638  *     (ebx+22) 16-bit cs  (this segment)
 639  *     (ebx+20) 16-bit ip  ('16-bit' return entry point)
 640  *     (ebx+16) 32-bit ss  (flat)
 641  *     (ebx+12) 32-bit sp  (32-bit stack pointer)
 642  *     (ebx+10) 16-bit bp  (points to ebx+24)
 643  *     (ebx+8)  16-bit si  (ignored)
 644  *     (ebx+6)  16-bit di  (ignored)
 645  *     (ebx+4)  16-bit ds  (we actually use the flat DS here)
 646  *     (ebx+2)  16-bit ss  (16-bit stack segment)
 647  *     (ebx+0)  16-bit sp  (points to ebx+4)
 648  *
 649  * Note that we ensure that DS is not changed and remains the flat segment,
 650  * and the 32-bit stack pointer our own return stub needs fits just
 651  * perfectly into the 8 bytes that are skipped by the Windows stub.
 652  * One problem is that we have to determine the number of removed arguments,
 653  * as these have to be really removed in KERNEL.621. Thus, the BP value
 654  * that we place in the temporary area to be restored, contains the value
 655  * that SP would have if no arguments were removed. By comparing the actual
 656  * value of SP with this value in our return stub we can compute the number
 657  * of removed arguments. This is then returned to KERNEL.621.
 658  *
 659  * The stack layout of this function:
 660  * (ebp+20)  nArgs     pointer to variable receiving nr. of args (Ex only)
 661  * (ebp+16)  esi       pointer to caller's esi value
 662  * (ebp+12)  arg       ebp value to be set for relay stub
 663  * (ebp+8)   func      CBClient relay stub address
 664  * (ebp+4)   ret addr
 665  * (ebp)     ebp
 666  */
 667 static void BuildCallTo32CBClient( FILE *outfile, BOOL isEx )
 668 {
 669     function_header( outfile, isEx ? "CALL32_CBClientEx" : "CALL32_CBClient" );
 670
 671     /* Entry code */
 672
 673     fprintf( outfile, "\tpushl %%ebp\n" );
 674     fprintf( outfile, "\tmovl %%esp,%%ebp\n" );
 675     fprintf( outfile, "\tpushl %%edi\n" );
 676     fprintf( outfile, "\tpushl %%esi\n" );
 677     fprintf( outfile, "\tpushl %%ebx\n" );
 678
 679     /* Get pointer to temporary area and save the 32-bit stack pointer */
 680
 681     fprintf( outfile, "\tmovl 16(%%ebp), %%ebx\n" );
 682     fprintf( outfile, "\tleal -8(%%esp), %%eax\n" );
 683
 684     if ( !isEx )
 685         fprintf( outfile, "\tmovl %%eax, -8(%%ebx)\n" );
 686     else
 687         fprintf( outfile, "\tmovl %%eax, 12(%%ebx)\n" );
 688
 689     /* Set up registers and call CBClient relay stub (simulating a far call) */
 690
 691     fprintf( outfile, "\tmovl 20(%%ebp), %%esi\n" );
 692     fprintf( outfile, "\tmovl (%%esi), %%esi\n" );
 693
 694     fprintf( outfile, "\tmovl 8(%%ebp), %%eax\n" );
 695     fprintf( outfile, "\tmovl 12(%%ebp), %%ebp\n" );
 696
 697     fprintf( outfile, "\tpushl %%cs\n" );
 698     fprintf( outfile, "\tcall *%%eax\n" );
 699
 700     /* Return new esi value to caller */
 701
 702     fprintf( outfile, "\tmovl 32(%%esp), %%edi\n" );
 703     fprintf( outfile, "\tmovl %%esi, (%%edi)\n" );
 704
 705     /* Return argument size to caller */
 706     if ( isEx )
 707     {
 708         fprintf( outfile, "\tmovl 36(%%esp), %%ebx\n" );
 709         fprintf( outfile, "\tmovl %%ebp, (%%ebx)\n" );
 710     }
 711
 712     /* Restore registers and return */
 713
 714     fprintf( outfile, "\tpopl %%ebx\n" );
 715     fprintf( outfile, "\tpopl %%esi\n" );
 716     fprintf( outfile, "\tpopl %%edi\n" );
 717     fprintf( outfile, "\tpopl %%ebp\n" );
 718     fprintf( outfile, "\tret\n" );
 719     output_function_size( outfile, isEx ? "CALL32_CBClientEx" : "CALL32_CBClient" );
 720
 721     /* '16-bit' return stub */
 722
 723     function_header( outfile, isEx ? "CALL32_CBClientEx_Ret" : "CALL32_CBClient_Ret" );
 724     if ( !isEx )
 725     {
 726         fprintf( outfile, "\tmovzwl %%sp, %%ebx\n" );
 727         fprintf( outfile, "\tlssl %%ss:-16(%%ebx), %%esp\n" );
 728     }
 729     else
 730     {
 731         fprintf( outfile, "\tmovzwl %%bp, %%ebx\n" );
 732         fprintf( outfile, "\tsubw %%bp, %%sp\n" );
 733         fprintf( outfile, "\tmovzwl %%sp, %%ebp\n" );
 734         fprintf( outfile, "\tlssl %%ss:-12(%%ebx), %%esp\n" );
 735     }
 736     fprintf( outfile, "\tlret\n" );
 737     output_function_size( outfile, isEx ? "CALL32_CBClientEx_Ret" : "CALL32_CBClient_Ret" );
 738 }
 739
 740
 741 /*******************************************************************
 742  *         BuildCallFrom32Regs
 743  *
 744  * Build a 32-bit-to-Wine call-back function for a 'register' function.
 745  * 'args' is the number of dword arguments.
 746  *
 747  * Stack layout:
 748  *   ...
 749  * (ebp+16)  first arg
 750  * (ebp+12)  ret addr to user code
 751  * (ebp+8)   eax saved by relay code
 752  * (ebp+4)   ret addr to relay code
 753  * (ebp+0)   saved ebp
 754  * (ebp-128) buffer area to allow stack frame manipulation
 755  * (ebp-332) CONTEXT86 struct
 756  * (ebp-336) padding for stack alignment
 757  * (ebp-336-n) CONTEXT86 *argument
 758  *  ....     other arguments copied from (ebp+12)
 759  *
 760  * The entry point routine is called with a CONTEXT* extra argument,
 761  * following the normal args. In this context structure, EIP_reg
 762  * contains the return address to user code, and ESP_reg the stack
 763  * pointer on return (with the return address and arguments already
 764  * removed).
 765  */
 766 static void BuildCallFrom32Regs( FILE *outfile )
 767 {
 768     static const int STACK_SPACE = 128 + sizeof(CONTEXT86);
 769
 770     /* Function header */
 771
 772     function_header( outfile, "__wine_call_from_32_regs" );
 773
 774     /* Allocate some buffer space on the stack */
 775
 776     fprintf( outfile, "\tpushl %%ebp\n" );
 777     fprintf( outfile, "\tmovl %%esp,%%ebp\n ");
 778     fprintf( outfile, "\tleal -%d(%%esp), %%esp\n", STACK_SPACE + 4 /* for context arg */);
 779
 780     /* Build the context structure */
 781
 782     fprintf( outfile, "\tpushfl\n" );
 783     fprintf( outfile, "\tpopl %%eax\n" );
 784     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(EFlags) - STACK_SPACE );
 785     fprintf( outfile, "\tmovl 0(%%ebp),%%eax\n" );
 786     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Ebp) - STACK_SPACE );
 787     fprintf( outfile, "\tmovl 8(%%ebp),%%eax\n" );
 788     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Eax) - STACK_SPACE );
 789     fprintf( outfile, "\tmovl %%ebx,%d(%%ebp)\n", CONTEXTOFFSET(Ebx) - STACK_SPACE );
 790     fprintf( outfile, "\tmovl %%ecx,%d(%%ebp)\n", CONTEXTOFFSET(Ecx) - STACK_SPACE );
 791     fprintf( outfile, "\tmovl %%edx,%d(%%ebp)\n", CONTEXTOFFSET(Edx) - STACK_SPACE );
 792     fprintf( outfile, "\tmovl %%esi,%d(%%ebp)\n", CONTEXTOFFSET(Esi) - STACK_SPACE );
 793     fprintf( outfile, "\tmovl %%edi,%d(%%ebp)\n", CONTEXTOFFSET(Edi) - STACK_SPACE );
 794
 795     fprintf( outfile, "\txorl %%eax,%%eax\n" );
 796     fprintf( outfile, "\tmovw %%cs,%%ax\n" );
 797     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegCs) - STACK_SPACE );
 798     fprintf( outfile, "\tmovw %%es,%%ax\n" );
 799     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegEs) - STACK_SPACE );
 800     fprintf( outfile, "\tmovw %%fs,%%ax\n" );
 801     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegFs) - STACK_SPACE );
 802     fprintf( outfile, "\tmovw %%gs,%%ax\n" );
 803     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegGs) - STACK_SPACE );
 804     fprintf( outfile, "\tmovw %%ss,%%ax\n" );
 805     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegSs) - STACK_SPACE );
 806     fprintf( outfile, "\tmovw %%ds,%%ax\n" );
 807     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegDs) - STACK_SPACE );
 808     fprintf( outfile, "\tmovw %%ax,%%es\n" );  /* set %es equal to %ds just in case */
 809
 810     fprintf( outfile, "\tmovl $0x%x,%%eax\n", CONTEXT86_FULL );
 811     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(ContextFlags) - STACK_SPACE );
 812
 813     fprintf( outfile, "\tmovl 12(%%ebp),%%eax\n" ); /* Get %eip at time of call */
 814     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Eip) - STACK_SPACE );
 815
 816     /* Transfer the arguments */
 817
 818     fprintf( outfile, "\tmovl 4(%%ebp),%%ebx\n" );   /* get relay code addr */
 819     fprintf( outfile, "\tmovzbl 4(%%ebx),%%ecx\n" ); /* fetch number of args to copy */
 820     fprintf( outfile, "\tsubl %%ecx,%%esp\n" );
 821     fprintf( outfile, "\tandl $~15,%%esp\n" );
 822     fprintf( outfile, "\tleal 16(%%ebp),%%esi\n" );  /* get %esp at time of call */
 823     fprintf( outfile, "\tmovl %%esp,%%edi\n" );
 824     fprintf( outfile, "\tshrl $2,%%ecx\n" );
 825     fprintf( outfile, "\tjz 1f\n" );
 826     fprintf( outfile, "\tcld\n" );
 827     fprintf( outfile, "\trep\n\tmovsl\n" );  /* copy args */
 828     fprintf( outfile, "1:\tleal %d(%%ebp),%%eax\n", -STACK_SPACE );  /* get addr of context struct */
 829     fprintf( outfile, "\tmovl %%eax,(%%edi)\n" );    /* and pass it as extra arg */
 830     fprintf( outfile, "\tmovzbl 5(%%ebx),%%eax\n" ); /* fetch number of args to remove */
 831     fprintf( outfile, "\tleal 16(%%ebp,%%eax),%%eax\n" );
 832     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Esp) - STACK_SPACE );
 833
 834     /* Call the entry point */
 835
 836     fprintf( outfile, "\taddl (%%ebx),%%ebx\n" );
 837     fprintf( outfile, "\tcall *%%ebx\n" );
 838     fprintf( outfile, "\tleal -%d(%%ebp),%%ecx\n", STACK_SPACE );
 839
 840     /* Restore the context structure */
 841
 842     fprintf( outfile, "2:\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegEs) );
 843     fprintf( outfile, "\tpopl %%es\n" );
 844     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegFs) );
 845     fprintf( outfile, "\tpopl %%fs\n" );
 846     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegGs) );
 847     fprintf( outfile, "\tpopl %%gs\n" );
 848
 849     fprintf( outfile, "\tmovl %d(%%ecx),%%edi\n", CONTEXTOFFSET(Edi) );
 850     fprintf( outfile, "\tmovl %d(%%ecx),%%esi\n", CONTEXTOFFSET(Esi) );
 851     fprintf( outfile, "\tmovl %d(%%ecx),%%edx\n", CONTEXTOFFSET(Edx) );
 852     fprintf( outfile, "\tmovl %d(%%ecx),%%ebx\n", CONTEXTOFFSET(Ebx) );
 853     fprintf( outfile, "\tmovl %d(%%ecx),%%eax\n", CONTEXTOFFSET(Eax) );
 854     fprintf( outfile, "\tmovl %d(%%ecx),%%ebp\n", CONTEXTOFFSET(Ebp) );
 855
 856     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegSs) );
 857     fprintf( outfile, "\tpopl %%ss\n" );
 858     fprintf( outfile, "\tmovl %d(%%ecx),%%esp\n", CONTEXTOFFSET(Esp) );
 859
 860     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(EFlags) );
 861     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegCs) );
 862     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(Eip) );
 863     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegDs) );
 864     fprintf( outfile, "\tmovl %d(%%ecx),%%ecx\n", CONTEXTOFFSET(Ecx) );
 865
 866     fprintf( outfile, "\tpopl %%ds\n" );
 867     fprintf( outfile, "\tiret\n" );
 868     output_function_size( outfile, "__wine_call_from_32_regs" );
 869
 870     function_header( outfile, "__wine_call_from_32_restore_regs" );
 871     fprintf( outfile, "\tmovl 4(%%esp),%%ecx\n" );
 872     fprintf( outfile, "\tjmp 2b\n" );
 873     output_function_size( outfile, "__wine_call_from_32_restore_regs" );
 874 }
 875
 876
 877 /*******************************************************************
 878  *         BuildPendingEventCheck
 879  *
 880  * Build a function that checks whether there are any
 881  * pending DPMI events.
 882  *
 883  * Stack layout:
 884  *
 885  * (sp+12) long   eflags
 886  * (sp+6)  long   cs
 887  * (sp+2)  long   ip
 888  * (sp)    word   fs
 889  *
 890  * On entry to function, fs register points to a valid TEB.
 891  * On exit from function, stack will be popped.
 892  */
 893 static void BuildPendingEventCheck( FILE *outfile )
 894 {
 895     /* Function header */
 896
 897     function_header( outfile, "DPMI_PendingEventCheck" );
 898
 899     /* Check for pending events. */
 900
 901     fprintf( outfile, "\t.byte 0x64\n\ttestl $0xffffffff,(%d)\n",
 902              STRUCTOFFSET(TEB,vm86_pending) );
 903     fprintf( outfile, "\tje %s\n", asm_name("DPMI_PendingEventCheck_Cleanup") );
 904
 905     fprintf( outfile, "\t.byte 0x64\n\ttestl $0xffffffff,(%d)\n",
 906              STRUCTOFFSET(TEB,dpmi_vif) );
 907
 908     fprintf( outfile, "\tje %s\n", asm_name("DPMI_PendingEventCheck_Cleanup") );
 909
 910     /* Process pending events. */
 911
 912     fprintf( outfile, "\tsti\n" );
 913
 914     /* Start cleanup. Restore fs register. */
 915
 916     fprintf( outfile, "%s\n", asm_globl("DPMI_PendingEventCheck_Cleanup") );
 917     fprintf( outfile, "\tpopw %%fs\n" );
 918
 919     /* Return from function. */
 920
 921     fprintf( outfile, "%s\n", asm_globl("DPMI_PendingEventCheck_Return") );
 922     fprintf( outfile, "\tiret\n" );
 923
 924     output_function_size( outfile, "DPMI_PendingEventCheck" );
 925 }
 926
 927
 928 /*******************************************************************
 929  *         BuildRelays16
 930  *
 931  * Build all the 16-bit relay callbacks
 932  */
 933 void BuildRelays16( FILE *outfile )
 934 {
 935     if (target_cpu != CPU_x86)
 936     {
 937         fprintf( outfile, "/* File not used with this architecture. Do not edit! */\n\n" );
 938         return;
 939     }
 940
 941     /* File header */
 942
 943     fprintf( outfile, "/* File generated automatically. Do not edit! */\n\n" );
 944     fprintf( outfile, "\t.text\n" );
 945
 946     fprintf( outfile, "%s:\n\n", asm_name("__wine_spec_thunk_text_16") );
 947
 948     fprintf( outfile, "%s\n", asm_globl("__wine_call16_start") );
 949
 950     /* Standard CallFrom16 routine */
 951     BuildCallFrom16Core( outfile, FALSE, FALSE );
 952
 953     /* Register CallFrom16 routine */
 954     BuildCallFrom16Core( outfile, TRUE, FALSE );
 955
 956     /* C16ThkSL CallFrom16 routine */
 957     BuildCallFrom16Core( outfile, FALSE, TRUE );
 958
 959     /* Standard CallTo16 routine */
 960     BuildCallTo16Core( outfile, 0 );
 961
 962     /* Register CallTo16 routine */
 963     BuildCallTo16Core( outfile, 1 );
 964
 965     /* Standard CallTo16 return stub */
 966     BuildRet16Func( outfile );
 967
 968     /* CBClientThunkSL routine */
 969     BuildCallTo32CBClient( outfile, FALSE );
 970
 971     /* CBClientThunkSLEx routine */
 972     BuildCallTo32CBClient( outfile, TRUE  );
 973
 974     /* Pending DPMI events check stub */
 975     BuildPendingEventCheck( outfile );
 976
 977     fprintf( outfile, "%s\n", asm_globl("__wine_call16_end") );
 978     output_function_size( outfile, "__wine_spec_thunk_text_16" );
 979
 980     /* Declare the return address and data selector variables */
 981     fprintf( outfile, "\n\t.data\n\t.align %d\n", get_alignment(4) );
 982     fprintf( outfile, "%s\n\t.long 0\n", asm_globl("CallTo16_DataSelector") );
 983     fprintf( outfile, "%s\n\t.long 0\n", asm_globl("CallTo16_TebSelector") );
 984     if (UsePIC) fprintf( outfile, "wine_ldt_copy_ptr:\t.long %s\n", asm_name("wine_ldt_copy") );
 985     output_gnu_stack_note( outfile );
 986 }
 987
 988 /*******************************************************************
 989  *         BuildRelays32
 990  *
 991  * Build all the 32-bit relay callbacks
 992  */
 993 void BuildRelays32( FILE *outfile )
 994 {
 995     if (target_cpu != CPU_x86)
 996     {
 997         fprintf( outfile, "/* File not used with this architecture. Do not edit! */\n\n" );
 998         return;
 999     }
1000
1001     /* File header */
1002
1003     fprintf( outfile, "/* File generated automatically. Do not edit! */\n\n" );
1004     fprintf( outfile, "\t.text\n" );
1005     fprintf( outfile, "%s:\n\n", asm_name("__wine_spec_thunk_text_32") );
1006
1007     /* 32-bit register entry point */
1008     BuildCallFrom32Regs( outfile );
1009
1010     output_function_size( outfile, "__wine_spec_thunk_text_32" );
1011     output_gnu_stack_note( outfile );
1012 }