tools/winebuild/relay.c

   1 /*
   2  * Relay calls helper routines
   3  *
   4  * Copyright 1993 Robert J. Amstadt
   5  * Copyright 1995 Martin von Loewis
   6  * Copyright 1995, 1996, 1997 Alexandre Julliard
   7  * Copyright 1997 Eric Youngdale
   8  * Copyright 1999 Ulrich Weigand
   9  *
  10  * This library is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2.1 of the License, or (at your option) any later version.
  14  *
  15  * This library is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * Lesser General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU Lesser General Public
  21  * License along with this library; if not, write to the Free Software
  22  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  23  */
  24
  25 #include "config.h"
  26 #include "wine/port.h"
  27
  28 #include <ctype.h>
  29
  30 #include "thread.h"
  31 #include "wine/winbase16.h"
  32
  33 #include "build.h"
  34
  35 static void function_header( FILE *outfile, const char *name )
  36 {
  37     fprintf( outfile, "\n\t.align %d\n", get_alignment(4) );
  38     fprintf( outfile, "\t%s\n", func_declaration(name) );
  39     fprintf( outfile, "%s\n", asm_globl(name) );
  40 }
  41
  42
  43 static inline const char *data16_prefix(void)
  44 {
  45     return (target_platform == PLATFORM_SVR4) ? "\tdata16\n" : "";
  46 }
  47
  48 /*******************************************************************
  49  *         BuildCallFrom16Core
  50  *
  51  * This routine builds the core routines used in 16->32 thunks:
  52  * CallFrom16Word, CallFrom16Long, CallFrom16Register, and CallFrom16Thunk.
  53  *
  54  * These routines are intended to be called via a far call (with 32-bit
  55  * operand size) from 16-bit code.  The 16-bit code stub must push %bp,
  56  * the 32-bit entry point to be called, and the argument conversion
  57  * routine to be used (see stack layout below).
  58  *
  59  * The core routine completes the STACK16FRAME on the 16-bit stack and
  60  * switches to the 32-bit stack.  Then, the argument conversion routine
  61  * is called; it gets passed the 32-bit entry point and a pointer to the
  62  * 16-bit arguments (on the 16-bit stack) as parameters. (You can either
  63  * use conversion routines automatically generated by BuildCallFrom16,
  64  * or write your own for special purposes.)
  65  *
  66  * The conversion routine must call the 32-bit entry point, passing it
  67  * the converted arguments, and return its return value to the core.
  68  * After the conversion routine has returned, the core switches back
  69  * to the 16-bit stack, converts the return value to the DX:AX format
  70  * (CallFrom16Long), and returns to the 16-bit call stub.  All parameters,
  71  * including %bp, are popped off the stack.
  72  *
  73  * The 16-bit call stub now returns to the caller, popping the 16-bit
  74  * arguments if necessary (pascal calling convention).
  75  *
  76  * In the case of a 'register' function, CallFrom16Register fills a
  77  * CONTEXT86 structure with the values all registers had at the point
  78  * the first instruction of the 16-bit call stub was about to be
  79  * executed.  A pointer to this CONTEXT86 is passed as third parameter
  80  * to the argument conversion routine, which typically passes it on
  81  * to the called 32-bit entry point.
  82  *
  83  * CallFrom16Thunk is a special variant used by the implementation of
  84  * the Win95 16->32 thunk functions C16ThkSL and C16ThkSL01 and is
  85  * implemented as follows:
  86  * On entry, the EBX register is set up to contain a flat pointer to the
  87  * 16-bit stack such that EBX+22 points to the first argument.
  88  * Then, the entry point is called, while EBP is set up to point
  89  * to the return address (on the 32-bit stack).
  90  * The called function returns with CX set to the number of bytes
  91  * to be popped of the caller's stack.
  92  *
  93  * Stack layout upon entry to the core routine (STACK16FRAME):
  94  *  ...           ...
  95  * (sp+24) word   first 16-bit arg
  96  * (sp+22) word   cs
  97  * (sp+20) word   ip
  98  * (sp+18) word   bp
  99  * (sp+14) long   32-bit entry point (reused for Win16 mutex recursion count)
 100  * (sp+12) word   ip of actual entry point (necessary for relay debugging)
 101  * (sp+8)  long   relay (argument conversion) function entry point
 102  * (sp+4)  long   cs of 16-bit entry point
 103  * (sp)    long   ip of 16-bit entry point
 104  *
 105  * Added on the stack:
 106  * (sp-2)  word   saved gs
 107  * (sp-4)  word   saved fs
 108  * (sp-6)  word   saved es
 109  * (sp-8)  word   saved ds
 110  * (sp-12) long   saved ebp
 111  * (sp-16) long   saved ecx
 112  * (sp-20) long   saved edx
 113  * (sp-24) long   saved previous stack
 114  */
 115 static void BuildCallFrom16Core( FILE *outfile, int reg_func, int thunk )
 116 {
 117     /* Function header */
 118     if (thunk) function_header( outfile, "__wine_call_from_16_thunk" );
 119     else if (reg_func) function_header( outfile, "__wine_call_from_16_regs" );
 120     else function_header( outfile, "__wine_call_from_16" );
 121
 122     /* Create STACK16FRAME (except STACK32FRAME link) */
 123     fprintf( outfile, "\tpushw %%gs\n" );
 124     fprintf( outfile, "\tpushw %%fs\n" );
 125     fprintf( outfile, "\tpushw %%es\n" );
 126     fprintf( outfile, "\tpushw %%ds\n" );
 127     fprintf( outfile, "\tpushl %%ebp\n" );
 128     fprintf( outfile, "\tpushl %%ecx\n" );
 129     fprintf( outfile, "\tpushl %%edx\n" );
 130
 131     /* Save original EFlags register */
 132     if (reg_func) fprintf( outfile, "\tpushfl\n" );
 133
 134     if ( UsePIC )
 135     {
 136         fprintf( outfile, "\tcall 1f\n" );
 137         fprintf( outfile, "1:\tpopl %%ecx\n" );
 138         fprintf( outfile, "\t.byte 0x2e\n\tmovl %s-1b(%%ecx),%%edx\n",
 139                  asm_name("CallTo16_DataSelector") );
 140     }
 141     else
 142         fprintf( outfile, "\t.byte 0x2e\n\tmovl %s,%%edx\n", asm_name("CallTo16_DataSelector") );
 143
 144     /* Load 32-bit segment registers */
 145     fprintf( outfile, "%s\tmovw %%dx, %%ds\n", data16_prefix() );
 146     fprintf( outfile, "%s\tmovw %%dx, %%es\n", data16_prefix() );
 147
 148     if ( UsePIC )
 149         fprintf( outfile, "\tmovw %s-1b(%%ecx), %%fs\n", asm_name("CallTo16_TebSelector") );
 150     else
 151         fprintf( outfile, "\tmovw %s, %%fs\n", asm_name("CallTo16_TebSelector") );
 152
 153     fprintf( outfile, "\t.byte 0x64\n\tmov (%d),%%gs\n", STRUCTOFFSET(TEB,gs_sel) );
 154
 155     /* Translate STACK16FRAME base to flat offset in %edx */
 156     fprintf( outfile, "\tmovw %%ss, %%dx\n" );
 157     fprintf( outfile, "\tandl $0xfff8, %%edx\n" );
 158     fprintf( outfile, "\tshrl $1, %%edx\n" );
 159     if (UsePIC)
 160     {
 161         fprintf( outfile, "\taddl wine_ldt_copy_ptr-1b(%%ecx),%%edx\n" );
 162         fprintf( outfile, "\tmovl (%%edx), %%edx\n" );
 163     }
 164     else
 165         fprintf( outfile, "\tmovl %s(%%edx), %%edx\n", asm_name("wine_ldt_copy") );
 166     fprintf( outfile, "\tmovzwl %%sp, %%ebp\n" );
 167     fprintf( outfile, "\tleal %d(%%ebp,%%edx), %%edx\n", reg_func ? 0 : -4 );
 168
 169     /* Get saved flags into %ecx */
 170     if (reg_func) fprintf( outfile, "\tpopl %%ecx\n" );
 171
 172     /* Get the 32-bit stack pointer from the TEB and complete STACK16FRAME */
 173     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d), %%ebp\n", STACKOFFSET );
 174     fprintf( outfile, "\tpushl %%ebp\n" );
 175
 176     /* Switch stacks */
 177     fprintf( outfile, "%s\t.byte 0x64\n\tmovw %%ss, (%d)\n", data16_prefix(), STACKOFFSET + 2 );
 178     fprintf( outfile, "\t.byte 0x64\n\tmovw %%sp, (%d)\n", STACKOFFSET );
 179     fprintf( outfile, "\tpushl %%ds\n" );
 180     fprintf( outfile, "\tpopl %%ss\n" );
 181     fprintf( outfile, "\tmovl %%ebp, %%esp\n" );
 182     fprintf( outfile, "\taddl $%d, %%ebp\n", STRUCTOFFSET(STACK32FRAME, ebp) );
 183
 184
 185     /* At this point:
 186        STACK16FRAME is completely set up
 187        DS, ES, SS: flat data segment
 188        FS: current TEB
 189        ESP: points to last STACK32FRAME
 190        EBP: points to ebp member of last STACK32FRAME
 191        EDX: points to current STACK16FRAME
 192        ECX: contains saved flags
 193        all other registers: unchanged */
 194
 195     /* Special case: C16ThkSL stub */
 196     if ( thunk )
 197     {
 198         /* Set up registers as expected and call thunk */
 199         fprintf( outfile, "\tleal %d(%%edx), %%ebx\n", sizeof(STACK16FRAME)-22 );
 200         fprintf( outfile, "\tleal -4(%%esp), %%ebp\n" );
 201
 202         fprintf( outfile, "\tcall *%d(%%edx)\n", STACK16OFFSET(entry_point) );
 203
 204         /* Switch stack back */
 205         fprintf( outfile, "\t.byte 0x64\n\tmovw (%d), %%ss\n", STACKOFFSET+2 );
 206         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%esp\n", STACKOFFSET );
 207         fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 208
 209         /* Restore registers and return directly to caller */
 210         fprintf( outfile, "\taddl $8, %%esp\n" );
 211         fprintf( outfile, "\tpopl %%ebp\n" );
 212         fprintf( outfile, "\tpopw %%ds\n" );
 213         fprintf( outfile, "\tpopw %%es\n" );
 214         fprintf( outfile, "\tpopw %%fs\n" );
 215         fprintf( outfile, "\tpopw %%gs\n" );
 216         fprintf( outfile, "\taddl $20, %%esp\n" );
 217
 218         fprintf( outfile, "\txorb %%ch, %%ch\n" );
 219         fprintf( outfile, "\tpopl %%ebx\n" );
 220         fprintf( outfile, "\taddw %%cx, %%sp\n" );
 221         fprintf( outfile, "\tpush %%ebx\n" );
 222
 223         fprintf( outfile, "\t.byte 0x66\n" );
 224         fprintf( outfile, "\tlret\n" );
 225
 226         return;
 227     }
 228
 229
 230     /* Build register CONTEXT */
 231     if ( reg_func )
 232     {
 233         fprintf( outfile, "\tsubl $%d, %%esp\n", sizeof(CONTEXT86) );
 234
 235         fprintf( outfile, "\tmovl %%ecx, %d(%%esp)\n", CONTEXTOFFSET(EFlags) );
 236
 237         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Eax) );
 238         fprintf( outfile, "\tmovl %%ebx, %d(%%esp)\n", CONTEXTOFFSET(Ebx) );
 239         fprintf( outfile, "\tmovl %%esi, %d(%%esp)\n", CONTEXTOFFSET(Esi) );
 240         fprintf( outfile, "\tmovl %%edi, %d(%%esp)\n", CONTEXTOFFSET(Edi) );
 241
 242         fprintf( outfile, "\tmovl %d(%%edx), %%eax\n", STACK16OFFSET(ebp) );
 243         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Ebp) );
 244         fprintf( outfile, "\tmovl %d(%%edx), %%eax\n", STACK16OFFSET(ecx) );
 245         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Ecx) );
 246         fprintf( outfile, "\tmovl %d(%%edx), %%eax\n", STACK16OFFSET(edx) );
 247         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Edx) );
 248
 249         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(ds) );
 250         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegDs) );
 251         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(es) );
 252         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegEs) );
 253         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(fs) );
 254         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegFs) );
 255         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(gs) );
 256         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegGs) );
 257
 258         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(cs) );
 259         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegCs) );
 260         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(ip) );
 261         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Eip) );
 262
 263         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%eax\n", STACKOFFSET+2 );
 264         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegSs) );
 265         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%eax\n", STACKOFFSET );
 266         fprintf( outfile, "\taddl $%d, %%eax\n", STACK16OFFSET(ip) );
 267         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Esp) );
 268 #if 0
 269         fprintf( outfile, "\tfsave %d(%%esp)\n", CONTEXTOFFSET(FloatSave) );
 270 #endif
 271
 272         /* Push address of CONTEXT86 structure -- popped by the relay routine */
 273         fprintf( outfile, "\tpushl %%esp\n" );
 274     }
 275
 276     /* Call relay routine (which will call the API entry point) */
 277     fprintf( outfile, "\tleal %d(%%edx), %%eax\n", sizeof(STACK16FRAME) );
 278     fprintf( outfile, "\tpushl %%eax\n" );
 279     fprintf( outfile, "\tpushl %d(%%edx)\n", STACK16OFFSET(entry_point) );
 280     fprintf( outfile, "\tcall *%d(%%edx)\n", STACK16OFFSET(relay) );
 281
 282     if ( reg_func )
 283     {
 284         fprintf( outfile, "\tleal -%d(%%ebp), %%ebx\n",
 285                  sizeof(CONTEXT) + STRUCTOFFSET(STACK32FRAME, ebp) );
 286
 287         /* Switch stack back */
 288         fprintf( outfile, "\t.byte 0x64\n\tmovw (%d), %%ss\n", STACKOFFSET+2 );
 289         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%esp\n", STACKOFFSET );
 290         fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 291
 292         /* Get return address to CallFrom16 stub */
 293         fprintf( outfile, "\taddw $%d, %%sp\n", STACK16OFFSET(callfrom_ip)-4 );
 294         fprintf( outfile, "\tpopl %%eax\n" );
 295         fprintf( outfile, "\tpopl %%edx\n" );
 296
 297         /* Restore all registers from CONTEXT */
 298         fprintf( outfile, "\tmovw %d(%%ebx), %%ss\n", CONTEXTOFFSET(SegSs) );
 299         fprintf( outfile, "\tmovl %d(%%ebx), %%esp\n", CONTEXTOFFSET(Esp) );
 300         fprintf( outfile, "\taddl $4, %%esp\n" );  /* room for final return address */
 301
 302         fprintf( outfile, "\tpushw %d(%%ebx)\n", CONTEXTOFFSET(SegCs) );
 303         fprintf( outfile, "\tpushw %d(%%ebx)\n", CONTEXTOFFSET(Eip) );
 304         fprintf( outfile, "\tpushl %%edx\n" );
 305         fprintf( outfile, "\tpushl %%eax\n" );
 306         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(EFlags) );
 307         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(SegDs) );
 308
 309         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(SegEs) );
 310         fprintf( outfile, "\tpopl %%es\n" );
 311         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(SegFs) );
 312         fprintf( outfile, "\tpopl %%fs\n" );
 313         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(SegGs) );
 314         fprintf( outfile, "\tpopl %%gs\n" );
 315
 316         fprintf( outfile, "\tmovl %d(%%ebx), %%ebp\n", CONTEXTOFFSET(Ebp) );
 317         fprintf( outfile, "\tmovl %d(%%ebx), %%esi\n", CONTEXTOFFSET(Esi) );
 318         fprintf( outfile, "\tmovl %d(%%ebx), %%edi\n", CONTEXTOFFSET(Edi) );
 319         fprintf( outfile, "\tmovl %d(%%ebx), %%eax\n", CONTEXTOFFSET(Eax) );
 320         fprintf( outfile, "\tmovl %d(%%ebx), %%edx\n", CONTEXTOFFSET(Edx) );
 321         fprintf( outfile, "\tmovl %d(%%ebx), %%ecx\n", CONTEXTOFFSET(Ecx) );
 322         fprintf( outfile, "\tmovl %d(%%ebx), %%ebx\n", CONTEXTOFFSET(Ebx) );
 323
 324         fprintf( outfile, "\tpopl %%ds\n" );
 325         fprintf( outfile, "\tpopfl\n" );
 326         fprintf( outfile, "\tlret\n" );
 327     }
 328     else
 329     {
 330         /* Switch stack back */
 331         fprintf( outfile, "\t.byte 0x64\n\tmovw (%d), %%ss\n", STACKOFFSET+2 );
 332         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%esp\n", STACKOFFSET );
 333         fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 334
 335         /* Restore registers */
 336         fprintf( outfile, "\tpopl %%edx\n" );
 337         fprintf( outfile, "\tpopl %%ecx\n" );
 338         fprintf( outfile, "\tpopl %%ebp\n" );
 339         fprintf( outfile, "\tpopw %%ds\n" );
 340         fprintf( outfile, "\tpopw %%es\n" );
 341         fprintf( outfile, "\tpopw %%fs\n" );
 342         fprintf( outfile, "\tpopw %%gs\n" );
 343
 344         /* Return to return stub which will return to caller */
 345         fprintf( outfile, "\tlret $12\n" );
 346     }
 347     if (thunk) output_function_size( outfile, "__wine_call_from_16_thunk" );
 348     else if (reg_func) output_function_size( outfile, "__wine_call_from_16_regs" );
 349     else output_function_size( outfile, "__wine_call_from_16" );
 350 }
 351
 352
 353 /*******************************************************************
 354  *         BuildCallTo16Core
 355  *
 356  * This routine builds the core routines used in 32->16 thunks:
 357  *
 358  * extern DWORD WINAPI wine_call_to_16( FARPROC16 target, DWORD cbArgs, PEXCEPTION_HANDLER handler );
 359  * extern void WINAPI wine_call_to_16_regs( CONTEXT86 *context, DWORD cbArgs, PEXCEPTION_HANDLER handler );
 360  *
 361  * These routines can be called directly from 32-bit code.
 362  *
 363  * All routines expect that the 16-bit stack contents (arguments) and the
 364  * return address (segptr to CallTo16_Ret) were already set up by the
 365  * caller; nb_args must contain the number of bytes to be conserved.  The
 366  * 16-bit SS:SP will be set accordinly.
 367  *
 368  * All other registers are either taken from the CONTEXT86 structure
 369  * or else set to default values.  The target routine address is either
 370  * given directly or taken from the CONTEXT86.
 371  */
 372 static void BuildCallTo16Core( FILE *outfile, int reg_func )
 373 {
 374     const char *name = reg_func ? "wine_call_to_16_regs" : "wine_call_to_16";
 375
 376     /* Function header */
 377     function_header( outfile, name );
 378
 379     /* Function entry sequence */
 380     fprintf( outfile, "\tpushl %%ebp\n" );
 381     fprintf( outfile, "\tmovl %%esp, %%ebp\n" );
 382
 383     /* Save the 32-bit registers */
 384     fprintf( outfile, "\tpushl %%ebx\n" );
 385     fprintf( outfile, "\tpushl %%esi\n" );
 386     fprintf( outfile, "\tpushl %%edi\n" );
 387     fprintf( outfile, "\t.byte 0x64\n\tmov %%gs,(%d)\n", STRUCTOFFSET(TEB,gs_sel) );
 388
 389     /* Setup exception frame */
 390     fprintf( outfile, "\t.byte 0x64\n\tpushl (%d)\n", STACKOFFSET );
 391     fprintf( outfile, "\tpushl 16(%%ebp)\n" ); /* handler */
 392     fprintf( outfile, "\t.byte 0x64\n\tpushl (%d)\n", STRUCTOFFSET(TEB,Tib.ExceptionList) );
 393     fprintf( outfile, "\t.byte 0x64\n\tmovl %%esp,(%d)\n", STRUCTOFFSET(TEB,Tib.ExceptionList) );
 394
 395     /* Call the actual CallTo16 routine (simulate a lcall) */
 396     fprintf( outfile, "\tpushl %%cs\n" );
 397     fprintf( outfile, "\tcall .L%s\n", name );
 398
 399     /* Remove exception frame */
 400     fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STRUCTOFFSET(TEB,Tib.ExceptionList) );
 401     fprintf( outfile, "\taddl $4, %%esp\n" );
 402     fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 403
 404     if ( !reg_func )
 405     {
 406         /* Convert return value */
 407         fprintf( outfile, "\tandl $0xffff,%%eax\n" );
 408         fprintf( outfile, "\tshll $16,%%edx\n" );
 409         fprintf( outfile, "\torl %%edx,%%eax\n" );
 410     }
 411     else
 412     {
 413         /*
 414          * Modify CONTEXT86 structure to contain new values
 415          *
 416          * NOTE:  We restore only EAX, EBX, EDX, EDX, EBP, and ESP.
 417          *        The segment registers as well as ESI and EDI should
 418          *        not be modified by a well-behaved 16-bit routine in
 419          *        any case.  [If necessary, we could restore them as well,
 420          *        at the cost of a somewhat less efficient return path.]
 421          */
 422
 423         fprintf( outfile, "\tmovl %d(%%esp), %%edi\n", STACK32OFFSET(target) - STACK32OFFSET(edi));
 424                 /* everything above edi has been popped already */
 425
 426         fprintf( outfile, "\tmovl %%eax, %d(%%edi)\n", CONTEXTOFFSET(Eax) );
 427         fprintf( outfile, "\tmovl %%ebx, %d(%%edi)\n", CONTEXTOFFSET(Ebx) );
 428         fprintf( outfile, "\tmovl %%ecx, %d(%%edi)\n", CONTEXTOFFSET(Ecx) );
 429         fprintf( outfile, "\tmovl %%edx, %d(%%edi)\n", CONTEXTOFFSET(Edx) );
 430         fprintf( outfile, "\tmovl %%ebp, %d(%%edi)\n", CONTEXTOFFSET(Ebp) );
 431         fprintf( outfile, "\tmovl %%esi, %d(%%edi)\n", CONTEXTOFFSET(Esp) );
 432                  /* The return glue code saved %esp into %esi */
 433     }
 434
 435     /* Restore the 32-bit registers */
 436     fprintf( outfile, "\tpopl %%edi\n" );
 437     fprintf( outfile, "\tpopl %%esi\n" );
 438     fprintf( outfile, "\tpopl %%ebx\n" );
 439
 440     /* Function exit sequence */
 441     fprintf( outfile, "\tpopl %%ebp\n" );
 442     fprintf( outfile, "\tret $12\n" );
 443
 444
 445     /* Start of the actual CallTo16 routine */
 446
 447     fprintf( outfile, ".L%s:\n", name );
 448
 449     /* Switch to the 16-bit stack */
 450     fprintf( outfile, "\tmovl %%esp,%%edx\n" );
 451     fprintf( outfile, "%s\t.byte 0x64\n\tmovw (%d),%%ss\n", data16_prefix(), STACKOFFSET + 2);
 452     fprintf( outfile, "\t.byte 0x64\n\tmovw (%d),%%sp\n", STACKOFFSET );
 453     fprintf( outfile, "\t.byte 0x64\n\tmovl %%edx,(%d)\n", STACKOFFSET );
 454
 455     /* Make %bp point to the previous stackframe (built by CallFrom16) */
 456     fprintf( outfile, "\tmovzwl %%sp,%%ebp\n" );
 457     fprintf( outfile, "\tleal %d(%%ebp),%%ebp\n", STACK16OFFSET(bp) );
 458
 459     /* Add the specified offset to the new sp */
 460     fprintf( outfile, "\tsubw %d(%%edx), %%sp\n", STACK32OFFSET(nb_args) );
 461
 462     if (reg_func)
 463     {
 464         /* Push the called routine address */
 465         fprintf( outfile, "\tmovl %d(%%edx),%%edx\n", STACK32OFFSET(target) );
 466         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(SegCs) );
 467         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(Eip) );
 468
 469         /* Get the registers */
 470         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(SegDs) );
 471         fprintf( outfile, "\tpushl %d(%%edx)\n", CONTEXTOFFSET(SegEs) );
 472         fprintf( outfile, "\tpopl %%es\n" );
 473         fprintf( outfile, "\tpushl %d(%%edx)\n", CONTEXTOFFSET(SegFs) );
 474         fprintf( outfile, "\tpopl %%fs\n" );
 475         fprintf( outfile, "\tpushl %d(%%edx)\n", CONTEXTOFFSET(SegGs) );
 476         fprintf( outfile, "\tpopl %%gs\n" );
 477         fprintf( outfile, "\tmovl %d(%%edx),%%ebp\n", CONTEXTOFFSET(Ebp) );
 478         fprintf( outfile, "\tmovl %d(%%edx),%%esi\n", CONTEXTOFFSET(Esi) );
 479         fprintf( outfile, "\tmovl %d(%%edx),%%edi\n", CONTEXTOFFSET(Edi) );
 480         fprintf( outfile, "\tmovl %d(%%edx),%%eax\n", CONTEXTOFFSET(Eax) );
 481         fprintf( outfile, "\tmovl %d(%%edx),%%ebx\n", CONTEXTOFFSET(Ebx) );
 482         fprintf( outfile, "\tmovl %d(%%edx),%%ecx\n", CONTEXTOFFSET(Ecx) );
 483         fprintf( outfile, "\tmovl %d(%%edx),%%edx\n", CONTEXTOFFSET(Edx) );
 484
 485         /* Get the 16-bit ds */
 486         fprintf( outfile, "\tpopw %%ds\n" );
 487     }
 488     else  /* not a register function */
 489     {
 490         /* Push the called routine address */
 491         fprintf( outfile, "\tpushl %d(%%edx)\n", STACK32OFFSET(target) );
 492
 493         /* Set %fs and %gs to the value saved by the last CallFrom16 */
 494         fprintf( outfile, "\tpushw %d(%%ebp)\n", STACK16OFFSET(fs)-STACK16OFFSET(bp) );
 495         fprintf( outfile, "\tpopw %%fs\n" );
 496         fprintf( outfile, "\tpushw %d(%%ebp)\n", STACK16OFFSET(gs)-STACK16OFFSET(bp) );
 497         fprintf( outfile, "\tpopw %%gs\n" );
 498
 499         /* Set %ds and %es (and %ax just in case) equal to %ss */
 500         fprintf( outfile, "\tmovw %%ss,%%ax\n" );
 501         fprintf( outfile, "\tmovw %%ax,%%ds\n" );
 502         fprintf( outfile, "\tmovw %%ax,%%es\n" );
 503     }
 504
 505     /* Jump to the called routine */
 506     fprintf( outfile, "\t.byte 0x66\n" );
 507     fprintf( outfile, "\tlret\n" );
 508
 509     /* Function footer */
 510     output_function_size( outfile, name );
 511 }
 512
 513
 514 /*******************************************************************
 515  *         BuildRet16Func
 516  *
 517  * Build the return code for 16-bit callbacks
 518  */
 519 static void BuildRet16Func( FILE *outfile )
 520 {
 521     function_header( outfile, "__wine_call_to_16_ret" );
 522
 523     /* Save %esp into %esi */
 524     fprintf( outfile, "\tmovl %%esp,%%esi\n" );
 525
 526     /* Restore 32-bit segment registers */
 527
 528     fprintf( outfile, "\t.byte 0x2e\n\tmovl %s", asm_name("CallTo16_DataSelector") );
 529     fprintf( outfile, "-%s,%%edi\n", asm_name("__wine_call16_start") );
 530     fprintf( outfile, "%s\tmovw %%di,%%ds\n", data16_prefix() );
 531     fprintf( outfile, "%s\tmovw %%di,%%es\n", data16_prefix() );
 532
 533     fprintf( outfile, "\t.byte 0x2e\n\tmov %s", asm_name("CallTo16_TebSelector") );
 534     fprintf( outfile, "-%s,%%fs\n", asm_name("__wine_call16_start") );
 535
 536     fprintf( outfile, "\t.byte 0x64\n\tmov (%d),%%gs\n", STRUCTOFFSET(TEB,gs_sel) );
 537
 538     /* Restore the 32-bit stack */
 539
 540     fprintf( outfile, "%s\tmovw %%di,%%ss\n", data16_prefix() );
 541     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d),%%esp\n", STACKOFFSET );
 542
 543     /* Return to caller */
 544
 545     fprintf( outfile, "\tlret\n" );
 546     output_function_size( outfile, "__wine_call_to_16_ret" );
 547 }
 548
 549
 550 /*******************************************************************
 551  *         BuildCallTo32CBClient
 552  *
 553  * Call a CBClient relay stub from 32-bit code (KERNEL.620).
 554  *
 555  * Since the relay stub is itself 32-bit, this should not be a problem;
 556  * unfortunately, the relay stubs are expected to switch back to a
 557  * 16-bit stack (and 16-bit code) after completion :-(
 558  *
 559  * This would conflict with our 16- vs. 32-bit stack handling, so
 560  * we simply switch *back* to our 32-bit stack before returning to
 561  * the caller ...
 562  *
 563  * The CBClient relay stub expects to be called with the following
 564  * 16-bit stack layout, and with ebp and ebx pointing into the 16-bit
 565  * stack at the designated places:
 566  *
 567  *    ...
 568  *  (ebp+14) original arguments to the callback routine
 569  *  (ebp+10) far return address to original caller
 570  *  (ebp+6)  Thunklet target address
 571  *  (ebp+2)  Thunklet relay ID code
 572  *  (ebp)    BP (saved by CBClientGlueSL)
 573  *  (ebp-2)  SI (saved by CBClientGlueSL)
 574  *  (ebp-4)  DI (saved by CBClientGlueSL)
 575  *  (ebp-6)  DS (saved by CBClientGlueSL)
 576  *
 577  *   ...     buffer space used by the 16-bit side glue for temp copies
 578  *
 579  *  (ebx+4)  far return address to 16-bit side glue code
 580  *  (ebx)    saved 16-bit ss:sp (pointing to ebx+4)
 581  *
 582  * The 32-bit side glue code accesses both the original arguments (via ebp)
 583  * and the temporary copies prepared by the 16-bit side glue (via ebx).
 584  * After completion, the stub will load ss:sp from the buffer at ebx
 585  * and perform a far return to 16-bit code.
 586  *
 587  * To trick the relay stub into returning to us, we replace the 16-bit
 588  * return address to the glue code by a cs:ip pair pointing to our
 589  * return entry point (the original return address is saved first).
 590  * Our return stub thus called will then reload the 32-bit ss:esp and
 591  * return to 32-bit code (by using and ss:esp value that we have also
 592  * pushed onto the 16-bit stack before and a cs:eip values found at
 593  * that position on the 32-bit stack).  The ss:esp to be restored is
 594  * found relative to the 16-bit stack pointer at:
 595  *
 596  *  (ebx-4)   ss  (flat)
 597  *  (ebx-8)   sp  (32-bit stack pointer)
 598  *
 599  * The second variant of this routine, CALL32_CBClientEx, which is used
 600  * to implement KERNEL.621, has to cope with yet another problem: Here,
 601  * the 32-bit side directly returns to the caller of the CBClient thunklet,
 602  * restoring registers saved by CBClientGlueSL and cleaning up the stack.
 603  * As we have to return to our 32-bit code first, we have to adapt the
 604  * layout of our temporary area so as to include values for the registers
 605  * that are to be restored, and later (in the implementation of KERNEL.621)
 606  * we *really* restore them. The return stub restores DS, DI, SI, and BP
 607  * from the stack, skips the next 8 bytes (CBClient relay code / target),
 608  * and then performs a lret NN, where NN is the number of arguments to be
 609  * removed. Thus, we prepare our temporary area as follows:
 610  *
 611  *     (ebx+22) 16-bit cs  (this segment)
 612  *     (ebx+20) 16-bit ip  ('16-bit' return entry point)
 613  *     (ebx+16) 32-bit ss  (flat)
 614  *     (ebx+12) 32-bit sp  (32-bit stack pointer)
 615  *     (ebx+10) 16-bit bp  (points to ebx+24)
 616  *     (ebx+8)  16-bit si  (ignored)
 617  *     (ebx+6)  16-bit di  (ignored)
 618  *     (ebx+4)  16-bit ds  (we actually use the flat DS here)
 619  *     (ebx+2)  16-bit ss  (16-bit stack segment)
 620  *     (ebx+0)  16-bit sp  (points to ebx+4)
 621  *
 622  * Note that we ensure that DS is not changed and remains the flat segment,
 623  * and the 32-bit stack pointer our own return stub needs fits just
 624  * perfectly into the 8 bytes that are skipped by the Windows stub.
 625  * One problem is that we have to determine the number of removed arguments,
 626  * as these have to be really removed in KERNEL.621. Thus, the BP value
 627  * that we place in the temporary area to be restored, contains the value
 628  * that SP would have if no arguments were removed. By comparing the actual
 629  * value of SP with this value in our return stub we can compute the number
 630  * of removed arguments. This is then returned to KERNEL.621.
 631  *
 632  * The stack layout of this function:
 633  * (ebp+20)  nArgs     pointer to variable receiving nr. of args (Ex only)
 634  * (ebp+16)  esi       pointer to caller's esi value
 635  * (ebp+12)  arg       ebp value to be set for relay stub
 636  * (ebp+8)   func      CBClient relay stub address
 637  * (ebp+4)   ret addr
 638  * (ebp)     ebp
 639  */
 640 static void BuildCallTo32CBClient( FILE *outfile, BOOL isEx )
 641 {
 642     function_header( outfile, isEx ? "CALL32_CBClientEx" : "CALL32_CBClient" );
 643
 644     /* Entry code */
 645
 646     fprintf( outfile, "\tpushl %%ebp\n" );
 647     fprintf( outfile, "\tmovl %%esp,%%ebp\n" );
 648     fprintf( outfile, "\tpushl %%edi\n" );
 649     fprintf( outfile, "\tpushl %%esi\n" );
 650     fprintf( outfile, "\tpushl %%ebx\n" );
 651
 652     /* Get pointer to temporary area and save the 32-bit stack pointer */
 653
 654     fprintf( outfile, "\tmovl 16(%%ebp), %%ebx\n" );
 655     fprintf( outfile, "\tleal -8(%%esp), %%eax\n" );
 656
 657     if ( !isEx )
 658         fprintf( outfile, "\tmovl %%eax, -8(%%ebx)\n" );
 659     else
 660         fprintf( outfile, "\tmovl %%eax, 12(%%ebx)\n" );
 661
 662     /* Set up registers and call CBClient relay stub (simulating a far call) */
 663
 664     fprintf( outfile, "\tmovl 20(%%ebp), %%esi\n" );
 665     fprintf( outfile, "\tmovl (%%esi), %%esi\n" );
 666
 667     fprintf( outfile, "\tmovl 8(%%ebp), %%eax\n" );
 668     fprintf( outfile, "\tmovl 12(%%ebp), %%ebp\n" );
 669
 670     fprintf( outfile, "\tpushl %%cs\n" );
 671     fprintf( outfile, "\tcall *%%eax\n" );
 672
 673     /* Return new esi value to caller */
 674
 675     fprintf( outfile, "\tmovl 32(%%esp), %%edi\n" );
 676     fprintf( outfile, "\tmovl %%esi, (%%edi)\n" );
 677
 678     /* Return argument size to caller */
 679     if ( isEx )
 680     {
 681         fprintf( outfile, "\tmovl 36(%%esp), %%ebx\n" );
 682         fprintf( outfile, "\tmovl %%ebp, (%%ebx)\n" );
 683     }
 684
 685     /* Restore registers and return */
 686
 687     fprintf( outfile, "\tpopl %%ebx\n" );
 688     fprintf( outfile, "\tpopl %%esi\n" );
 689     fprintf( outfile, "\tpopl %%edi\n" );
 690     fprintf( outfile, "\tpopl %%ebp\n" );
 691     fprintf( outfile, "\tret\n" );
 692     output_function_size( outfile, isEx ? "CALL32_CBClientEx" : "CALL32_CBClient" );
 693
 694     /* '16-bit' return stub */
 695
 696     function_header( outfile, isEx ? "CALL32_CBClientEx_Ret" : "CALL32_CBClient_Ret" );
 697     if ( !isEx )
 698     {
 699         fprintf( outfile, "\tmovzwl %%sp, %%ebx\n" );
 700         fprintf( outfile, "\tlssl %%ss:-16(%%ebx), %%esp\n" );
 701     }
 702     else
 703     {
 704         fprintf( outfile, "\tmovzwl %%bp, %%ebx\n" );
 705         fprintf( outfile, "\tsubw %%bp, %%sp\n" );
 706         fprintf( outfile, "\tmovzwl %%sp, %%ebp\n" );
 707         fprintf( outfile, "\tlssl %%ss:-12(%%ebx), %%esp\n" );
 708     }
 709     fprintf( outfile, "\tlret\n" );
 710     output_function_size( outfile, isEx ? "CALL32_CBClientEx_Ret" : "CALL32_CBClient_Ret" );
 711 }
 712
 713
 714 /*******************************************************************
 715  *         BuildCallFrom32Regs
 716  *
 717  * Build a 32-bit-to-Wine call-back function for a 'register' function.
 718  * 'args' is the number of dword arguments.
 719  *
 720  * Stack layout:
 721  *   ...
 722  * (ebp+16)  first arg
 723  * (ebp+12)  ret addr to user code
 724  * (ebp+8)   eax saved by relay code
 725  * (ebp+4)   ret addr to relay code
 726  * (ebp+0)   saved ebp
 727  * (ebp-128) buffer area to allow stack frame manipulation
 728  * (ebp-332) CONTEXT86 struct
 729  * (ebp-336) CONTEXT86 *argument
 730  *  ....     other arguments copied from (ebp+12)
 731  *
 732  * The entry point routine is called with a CONTEXT* extra argument,
 733  * following the normal args. In this context structure, EIP_reg
 734  * contains the return address to user code, and ESP_reg the stack
 735  * pointer on return (with the return address and arguments already
 736  * removed).
 737  */
 738 static void BuildCallFrom32Regs( FILE *outfile )
 739 {
 740     static const int STACK_SPACE = 128 + sizeof(CONTEXT86);
 741
 742     /* Function header */
 743
 744     function_header( outfile, "__wine_call_from_32_regs" );
 745
 746     /* Allocate some buffer space on the stack */
 747
 748     fprintf( outfile, "\tpushl %%ebp\n" );
 749     fprintf( outfile, "\tmovl %%esp,%%ebp\n ");
 750     fprintf( outfile, "\tleal -%d(%%esp), %%esp\n", STACK_SPACE );
 751
 752     /* Build the context structure */
 753
 754     fprintf( outfile, "\tpushfl\n" );
 755     fprintf( outfile, "\tpopl %%eax\n" );
 756     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(EFlags) - STACK_SPACE );
 757     fprintf( outfile, "\tmovl 0(%%ebp),%%eax\n" );
 758     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Ebp) - STACK_SPACE );
 759     fprintf( outfile, "\tmovl 8(%%ebp),%%eax\n" );
 760     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Eax) - STACK_SPACE );
 761     fprintf( outfile, "\tmovl %%ebx,%d(%%ebp)\n", CONTEXTOFFSET(Ebx) - STACK_SPACE );
 762     fprintf( outfile, "\tmovl %%ecx,%d(%%ebp)\n", CONTEXTOFFSET(Ecx) - STACK_SPACE );
 763     fprintf( outfile, "\tmovl %%edx,%d(%%ebp)\n", CONTEXTOFFSET(Edx) - STACK_SPACE );
 764     fprintf( outfile, "\tmovl %%esi,%d(%%ebp)\n", CONTEXTOFFSET(Esi) - STACK_SPACE );
 765     fprintf( outfile, "\tmovl %%edi,%d(%%ebp)\n", CONTEXTOFFSET(Edi) - STACK_SPACE );
 766
 767     fprintf( outfile, "\txorl %%eax,%%eax\n" );
 768     fprintf( outfile, "\tmovw %%cs,%%ax\n" );
 769     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegCs) - STACK_SPACE );
 770     fprintf( outfile, "\tmovw %%es,%%ax\n" );
 771     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegEs) - STACK_SPACE );
 772     fprintf( outfile, "\tmovw %%fs,%%ax\n" );
 773     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegFs) - STACK_SPACE );
 774     fprintf( outfile, "\tmovw %%gs,%%ax\n" );
 775     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegGs) - STACK_SPACE );
 776     fprintf( outfile, "\tmovw %%ss,%%ax\n" );
 777     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegSs) - STACK_SPACE );
 778     fprintf( outfile, "\tmovw %%ds,%%ax\n" );
 779     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegDs) - STACK_SPACE );
 780     fprintf( outfile, "\tmovw %%ax,%%es\n" );  /* set %es equal to %ds just in case */
 781
 782     fprintf( outfile, "\tmovl $0x%x,%%eax\n", CONTEXT86_FULL );
 783     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(ContextFlags) - STACK_SPACE );
 784
 785     fprintf( outfile, "\tmovl 12(%%ebp),%%eax\n" ); /* Get %eip at time of call */
 786     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Eip) - STACK_SPACE );
 787
 788     /* Transfer the arguments */
 789
 790     fprintf( outfile, "\tmovl 4(%%ebp),%%ebx\n" );   /* get relay code addr */
 791     fprintf( outfile, "\tpushl %%esp\n" );           /* push ptr to context struct */
 792     fprintf( outfile, "\tmovzbl 4(%%ebx),%%ecx\n" ); /* fetch number of args to copy */
 793     fprintf( outfile, "\tjecxz 1f\n" );
 794     fprintf( outfile, "\tsubl %%ecx,%%esp\n" );
 795     fprintf( outfile, "\tleal 16(%%ebp),%%esi\n" );  /* get %esp at time of call */
 796     fprintf( outfile, "\tmovl %%esp,%%edi\n" );
 797     fprintf( outfile, "\tshrl $2,%%ecx\n" );
 798     fprintf( outfile, "\tcld\n" );
 799     fprintf( outfile, "\trep\n\tmovsl\n" );  /* copy args */
 800
 801     fprintf( outfile, "1:\tmovzbl 5(%%ebx),%%eax\n" ); /* fetch number of args to remove */
 802     fprintf( outfile, "\tleal 16(%%ebp,%%eax),%%eax\n" );
 803     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Esp) - STACK_SPACE );
 804
 805     /* Call the entry point */
 806
 807     fprintf( outfile, "\taddl (%%ebx),%%ebx\n" );
 808     fprintf( outfile, "\tcall *%%ebx\n" );
 809     fprintf( outfile, "\tleal -%d(%%ebp),%%ecx\n", STACK_SPACE );
 810
 811     /* Restore the context structure */
 812
 813     fprintf( outfile, "2:\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegEs) );
 814     fprintf( outfile, "\tpopl %%es\n" );
 815     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegFs) );
 816     fprintf( outfile, "\tpopl %%fs\n" );
 817     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegGs) );
 818     fprintf( outfile, "\tpopl %%gs\n" );
 819
 820     fprintf( outfile, "\tmovl %d(%%ecx),%%edi\n", CONTEXTOFFSET(Edi) );
 821     fprintf( outfile, "\tmovl %d(%%ecx),%%esi\n", CONTEXTOFFSET(Esi) );
 822     fprintf( outfile, "\tmovl %d(%%ecx),%%edx\n", CONTEXTOFFSET(Edx) );
 823     fprintf( outfile, "\tmovl %d(%%ecx),%%ebx\n", CONTEXTOFFSET(Ebx) );
 824     fprintf( outfile, "\tmovl %d(%%ecx),%%eax\n", CONTEXTOFFSET(Eax) );
 825     fprintf( outfile, "\tmovl %d(%%ecx),%%ebp\n", CONTEXTOFFSET(Ebp) );
 826
 827     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegSs) );
 828     fprintf( outfile, "\tpopl %%ss\n" );
 829     fprintf( outfile, "\tmovl %d(%%ecx),%%esp\n", CONTEXTOFFSET(Esp) );
 830
 831     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(EFlags) );
 832     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegCs) );
 833     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(Eip) );
 834     fprintf( outfile, "\tpushl %d(%%ecx)\n", CONTEXTOFFSET(SegDs) );
 835     fprintf( outfile, "\tmovl %d(%%ecx),%%ecx\n", CONTEXTOFFSET(Ecx) );
 836
 837     fprintf( outfile, "\tpopl %%ds\n" );
 838     fprintf( outfile, "\tiret\n" );
 839     output_function_size( outfile, "__wine_call_from_32_regs" );
 840
 841     function_header( outfile, "__wine_call_from_32_restore_regs" );
 842     fprintf( outfile, "\tleal 4(%%esp),%%ecx\n" );
 843     fprintf( outfile, "\tjmp 2b\n" );
 844     output_function_size( outfile, "__wine_call_from_32_restore_regs" );
 845 }
 846
 847
 848 /*******************************************************************
 849  *         BuildPendingEventCheck
 850  *
 851  * Build a function that checks whether there are any
 852  * pending DPMI events.
 853  *
 854  * Stack layout:
 855  *
 856  * (sp+12) long   eflags
 857  * (sp+6)  long   cs
 858  * (sp+2)  long   ip
 859  * (sp)    word   fs
 860  *
 861  * On entry to function, fs register points to a valid TEB.
 862  * On exit from function, stack will be popped.
 863  */
 864 static void BuildPendingEventCheck( FILE *outfile )
 865 {
 866     /* Function header */
 867
 868     function_header( outfile, "DPMI_PendingEventCheck" );
 869
 870     /* Check for pending events. */
 871
 872     fprintf( outfile, "\t.byte 0x64\n\ttestl $0xffffffff,(%d)\n",
 873              STRUCTOFFSET(TEB,vm86_pending) );
 874     fprintf( outfile, "\tje %s\n", asm_name("DPMI_PendingEventCheck_Cleanup") );
 875
 876     fprintf( outfile, "\t.byte 0x64\n\ttestl $0xffffffff,(%d)\n",
 877              STRUCTOFFSET(TEB,dpmi_vif) );
 878
 879     fprintf( outfile, "\tje %s\n", asm_name("DPMI_PendingEventCheck_Cleanup") );
 880
 881     /* Process pending events. */
 882
 883     fprintf( outfile, "\tsti\n" );
 884
 885     /* Start cleanup. Restore fs register. */
 886
 887     fprintf( outfile, "%s\n", asm_globl("DPMI_PendingEventCheck_Cleanup") );
 888     fprintf( outfile, "\tpopw %%fs\n" );
 889
 890     /* Return from function. */
 891
 892     fprintf( outfile, "%s\n", asm_globl("DPMI_PendingEventCheck_Return") );
 893     fprintf( outfile, "\tiret\n" );
 894
 895     output_function_size( outfile, "DPMI_PendingEventCheck" );
 896 }
 897
 898
 899 /*******************************************************************
 900  *         BuildRelays16
 901  *
 902  * Build all the 16-bit relay callbacks
 903  */
 904 void BuildRelays16( FILE *outfile )
 905 {
 906     if (target_cpu != CPU_x86)
 907     {
 908         fprintf( outfile, "/* File not used with this architecture. Do not edit! */\n\n" );
 909         return;
 910     }
 911
 912     /* File header */
 913
 914     fprintf( outfile, "/* File generated automatically. Do not edit! */\n\n" );
 915     fprintf( outfile, "\t.text\n" );
 916
 917     fprintf( outfile, "%s:\n\n", asm_name("__wine_spec_thunk_text_16") );
 918
 919     fprintf( outfile, "%s\n", asm_globl("__wine_call16_start") );
 920
 921     /* Standard CallFrom16 routine */
 922     BuildCallFrom16Core( outfile, FALSE, FALSE );
 923
 924     /* Register CallFrom16 routine */
 925     BuildCallFrom16Core( outfile, TRUE, FALSE );
 926
 927     /* C16ThkSL CallFrom16 routine */
 928     BuildCallFrom16Core( outfile, FALSE, TRUE );
 929
 930     /* Standard CallTo16 routine */
 931     BuildCallTo16Core( outfile, 0 );
 932
 933     /* Register CallTo16 routine */
 934     BuildCallTo16Core( outfile, 1 );
 935
 936     /* Standard CallTo16 return stub */
 937     BuildRet16Func( outfile );
 938
 939     /* CBClientThunkSL routine */
 940     BuildCallTo32CBClient( outfile, FALSE );
 941
 942     /* CBClientThunkSLEx routine */
 943     BuildCallTo32CBClient( outfile, TRUE  );
 944
 945     /* Pending DPMI events check stub */
 946     BuildPendingEventCheck( outfile );
 947
 948     fprintf( outfile, "%s\n", asm_globl("__wine_call16_end") );
 949     output_function_size( outfile, "__wine_spec_thunk_text_16" );
 950
 951     /* Declare the return address and data selector variables */
 952     fprintf( outfile, "\n\t.data\n\t.align %d\n", get_alignment(4) );
 953     fprintf( outfile, "%s\n\t.long 0\n", asm_globl("CallTo16_DataSelector") );
 954     fprintf( outfile, "%s\n\t.long 0\n", asm_globl("CallTo16_TebSelector") );
 955     if (UsePIC) fprintf( outfile, "wine_ldt_copy_ptr:\t.long %s\n", asm_name("wine_ldt_copy") );
 956 }
 957
 958 /*******************************************************************
 959  *         BuildRelays32
 960  *
 961  * Build all the 32-bit relay callbacks
 962  */
 963 void BuildRelays32( FILE *outfile )
 964 {
 965     if (target_cpu != CPU_x86)
 966     {
 967         fprintf( outfile, "/* File not used with this architecture. Do not edit! */\n\n" );
 968         return;
 969     }
 970
 971     /* File header */
 972
 973     fprintf( outfile, "/* File generated automatically. Do not edit! */\n\n" );
 974     fprintf( outfile, "\t.text\n" );
 975     fprintf( outfile, "%s:\n\n", asm_name("__wine_spec_thunk_text_32") );
 976
 977     /* 32-bit register entry point */
 978     BuildCallFrom32Regs( outfile );
 979
 980     output_function_size( outfile, "__wine_spec_thunk_text_32" );
 981 }