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
  27 #include <ctype.h>
  28 #include <stdarg.h>
  29
  30 #include "build.h"
  31
  32 /* offset of the stack pointer relative to %fs:(0) */
  33 #define STACKOFFSET 0x10c  /* FIELD_OFFSET(TEB,SystemReserved1) */
  34
  35 /* fix this if the x86_thread_data structure is changed */
  36 #define GS_OFFSET  0x1d8  /* FIELD_OFFSET(TEB,SystemReserved2) + FIELD_OFFSET(struct x86_thread_data,gs) */
  37
  38
  39 /*******************************************************************
  40  *         BuildCallFrom16Core
  41  *
  42  * This routine builds the core routines used in 16->32 thunks:
  43  * CallFrom16Word, CallFrom16Long, CallFrom16Register, and CallFrom16Thunk.
  44  *
  45  * These routines are intended to be called via a far call (with 32-bit
  46  * operand size) from 16-bit code.  The 16-bit code stub must push %bp,
  47  * the 32-bit entry point to be called, and the argument conversion
  48  * routine to be used (see stack layout below).
  49  *
  50  * The core routine completes the STACK16FRAME on the 16-bit stack and
  51  * switches to the 32-bit stack.  Then, the argument conversion routine
  52  * is called; it gets passed the 32-bit entry point and a pointer to the
  53  * 16-bit arguments (on the 16-bit stack) as parameters. (You can either
  54  * use conversion routines automatically generated by BuildCallFrom16,
  55  * or write your own for special purposes.)
  56  *
  57  * The conversion routine must call the 32-bit entry point, passing it
  58  * the converted arguments, and return its return value to the core.
  59  * After the conversion routine has returned, the core switches back
  60  * to the 16-bit stack, converts the return value to the DX:AX format
  61  * (CallFrom16Long), and returns to the 16-bit call stub.  All parameters,
  62  * including %bp, are popped off the stack.
  63  *
  64  * The 16-bit call stub now returns to the caller, popping the 16-bit
  65  * arguments if necessary (pascal calling convention).
  66  *
  67  * In the case of a 'register' function, CallFrom16Register fills a
  68  * CONTEXT86 structure with the values all registers had at the point
  69  * the first instruction of the 16-bit call stub was about to be
  70  * executed.  A pointer to this CONTEXT86 is passed as third parameter
  71  * to the argument conversion routine, which typically passes it on
  72  * to the called 32-bit entry point.
  73  *
  74  * CallFrom16Thunk is a special variant used by the implementation of
  75  * the Win95 16->32 thunk functions C16ThkSL and C16ThkSL01 and is
  76  * implemented as follows:
  77  * On entry, the EBX register is set up to contain a flat pointer to the
  78  * 16-bit stack such that EBX+22 points to the first argument.
  79  * Then, the entry point is called, while EBP is set up to point
  80  * to the return address (on the 32-bit stack).
  81  * The called function returns with CX set to the number of bytes
  82  * to be popped of the caller's stack.
  83  *
  84  * Stack layout upon entry to the core routine (STACK16FRAME):
  85  *  ...           ...
  86  * (sp+24) word   first 16-bit arg
  87  * (sp+22) word   cs
  88  * (sp+20) word   ip
  89  * (sp+18) word   bp
  90  * (sp+14) long   32-bit entry point (reused for Win16 mutex recursion count)
  91  * (sp+12) word   ip of actual entry point (necessary for relay debugging)
  92  * (sp+8)  long   relay (argument conversion) function entry point
  93  * (sp+4)  long   cs of 16-bit entry point
  94  * (sp)    long   ip of 16-bit entry point
  95  *
  96  * Added on the stack:
  97  * (sp-2)  word   saved gs
  98  * (sp-4)  word   saved fs
  99  * (sp-6)  word   saved es
 100  * (sp-8)  word   saved ds
 101  * (sp-12) long   saved ebp
 102  * (sp-16) long   saved ecx
 103  * (sp-20) long   saved edx
 104  * (sp-24) long   saved previous stack
 105  */
 106 static void BuildCallFrom16Core( int reg_func, int thunk )
 107 {
 108     /* Function header */
 109     if (thunk) output_function_header( "__wine_call_from_16_thunk", 1 );
 110     else if (reg_func) output_function_header( "__wine_call_from_16_regs", 1 );
 111     else output_function_header( "__wine_call_from_16", 1 );
 112
 113     /* Create STACK16FRAME (except STACK32FRAME link) */
 114     output( "\tpushw %%gs\n" );
 115     output( "\tpushw %%fs\n" );
 116     output( "\tpushw %%es\n" );
 117     output( "\tpushw %%ds\n" );
 118     output( "\tpushl %%ebp\n" );
 119     output( "\tpushl %%ecx\n" );
 120     output( "\tpushl %%edx\n" );
 121
 122     /* Save original EFlags register */
 123     if (reg_func) output( "\tpushfl\n" );
 124
 125     if ( UsePIC )
 126     {
 127         output( "\tcall 1f\n" );
 128         output( "1:\tpopl %%ecx\n" );
 129         output( "\tmovl %%cs:%s-1b(%%ecx),%%edx\n", asm_name("CallTo16_DataSelector") );
 130     }
 131     else
 132         output( "\tmovl %%cs:%s,%%edx\n", asm_name("CallTo16_DataSelector") );
 133
 134     /* Load 32-bit segment registers */
 135     output( "\tmovw %%dx, %%ds\n" );
 136     output( "\tmovw %%dx, %%es\n" );
 137
 138     if ( UsePIC )
 139         output( "\tmovw %s-1b(%%ecx), %%fs\n", asm_name("CallTo16_TebSelector") );
 140     else
 141         output( "\tmovw %s, %%fs\n", asm_name("CallTo16_TebSelector") );
 142
 143     output( "\tmov %%fs:(%d),%%gs\n", GS_OFFSET );
 144
 145     /* Translate STACK16FRAME base to flat offset in %edx */
 146     output( "\tmovw %%ss, %%dx\n" );
 147     output( "\tandl $0xfff8, %%edx\n" );
 148     output( "\tshrl $1, %%edx\n" );
 149     if (UsePIC)
 150         output( "\taddl .Lwine_ldt_copy_ptr-1b(%%ecx),%%edx\n" );
 151     else
 152         output( "\taddl .Lwine_ldt_copy_ptr,%%edx\n" );
 153     output( "\tmovl (%%edx), %%edx\n" );
 154     output( "\tmovzwl %%sp, %%ebp\n" );
 155     output( "\tleal %d(%%ebp,%%edx), %%edx\n", reg_func ? 0 : -4 );
 156
 157     /* Get saved flags into %ecx */
 158     if (reg_func) output( "\tpopl %%ecx\n" );
 159
 160     /* Get the 32-bit stack pointer from the TEB and complete STACK16FRAME */
 161     output( "\tmovl %%fs:(%d), %%ebp\n", STACKOFFSET );
 162     output( "\tpushl %%ebp\n" );
 163
 164     /* Switch stacks */
 165     output( "\tmovw %%ss, %%fs:(%d)\n", STACKOFFSET + 2 );
 166     output( "\tmovw %%sp, %%fs:(%d)\n", STACKOFFSET );
 167     output( "\tpushl %%ds\n" );
 168     output( "\tpopl %%ss\n" );
 169     output( "\tmovl %%ebp, %%esp\n" );
 170     output( "\taddl $0x20,%%ebp\n");  /* FIELD_OFFSET(STACK32FRAME,ebp) */
 171
 172
 173     /* At this point:
 174        STACK16FRAME is completely set up
 175        DS, ES, SS: flat data segment
 176        FS: current TEB
 177        ESP: points to last STACK32FRAME
 178        EBP: points to ebp member of last STACK32FRAME
 179        EDX: points to current STACK16FRAME
 180        ECX: contains saved flags
 181        all other registers: unchanged */
 182
 183     /* Special case: C16ThkSL stub */
 184     if ( thunk )
 185     {
 186         /* Set up registers as expected and call thunk */
 187         output( "\tleal 0x1a(%%edx),%%ebx\n" );  /* sizeof(STACK16FRAME)-22 */
 188         output( "\tleal -4(%%esp), %%ebp\n" );
 189
 190         output( "\tcall *0x26(%%edx)\n");  /* FIELD_OFFSET(STACK16FRAME,entry_point) */
 191
 192         /* Switch stack back */
 193         output( "\tmovw %%fs:(%d), %%ss\n", STACKOFFSET+2 );
 194         output( "\tmovzwl %%fs:(%d), %%esp\n", STACKOFFSET );
 195         output( "\tpopl %%fs:(%d)\n", STACKOFFSET );
 196
 197         /* Restore registers and return directly to caller */
 198         output( "\taddl $8, %%esp\n" );
 199         output( "\tpopl %%ebp\n" );
 200         output( "\tpopw %%ds\n" );
 201         output( "\tpopw %%es\n" );
 202         output( "\tpopw %%fs\n" );
 203         output( "\tpopw %%gs\n" );
 204         output( "\taddl $20, %%esp\n" );
 205
 206         output( "\txorb %%ch, %%ch\n" );
 207         output( "\tpopl %%ebx\n" );
 208         output( "\taddw %%cx, %%sp\n" );
 209         output( "\tpush %%ebx\n" );
 210
 211         output( "\t.byte 0x66\n" );
 212         output( "\tlret\n" );
 213
 214         output_function_size( "__wine_call_from_16_thunk" );
 215         return;
 216     }
 217
 218
 219     /* Build register CONTEXT */
 220     if ( reg_func )
 221     {
 222         output( "\tsubl $0x2cc,%%esp\n" );       /* sizeof(CONTEXT86) */
 223
 224         output( "\tmovl %%ecx,0xc0(%%esp)\n" );  /* EFlags */
 225
 226         output( "\tmovl %%eax,0xb0(%%esp)\n" );  /* Eax */
 227         output( "\tmovl %%ebx,0xa4(%%esp)\n" );  /* Ebx */
 228         output( "\tmovl %%esi,0xa0(%%esp)\n" );  /* Esi */
 229         output( "\tmovl %%edi,0x9c(%%esp)\n" );  /* Edi */
 230
 231         output( "\tmovl 0x0c(%%edx),%%eax\n");   /* FIELD_OFFSET(STACK16FRAME,ebp) */
 232         output( "\tmovl %%eax,0xb4(%%esp)\n" );  /* Ebp */
 233         output( "\tmovl 0x08(%%edx),%%eax\n");   /* FIELD_OFFSET(STACK16FRAME,ecx) */
 234         output( "\tmovl %%eax,0xac(%%esp)\n" );  /* Ecx */
 235         output( "\tmovl 0x04(%%edx),%%eax\n");   /* FIELD_OFFSET(STACK16FRAME,edx) */
 236         output( "\tmovl %%eax,0xa8(%%esp)\n" );  /* Edx */
 237
 238         output( "\tmovzwl 0x10(%%edx),%%eax\n"); /* FIELD_OFFSET(STACK16FRAME,ds) */
 239         output( "\tmovl %%eax,0x98(%%esp)\n" );  /* SegDs */
 240         output( "\tmovzwl 0x12(%%edx),%%eax\n"); /* FIELD_OFFSET(STACK16FRAME,es) */
 241         output( "\tmovl %%eax,0x94(%%esp)\n" );  /* SegEs */
 242         output( "\tmovzwl 0x14(%%edx),%%eax\n"); /* FIELD_OFFSET(STACK16FRAME,fs) */
 243         output( "\tmovl %%eax,0x90(%%esp)\n" );  /* SegFs */
 244         output( "\tmovzwl 0x16(%%edx),%%eax\n"); /* FIELD_OFFSET(STACK16FRAME,gs) */
 245         output( "\tmovl %%eax,0x8c(%%esp)\n" );  /* SegGs */
 246
 247         output( "\tmovzwl 0x2e(%%edx),%%eax\n"); /* FIELD_OFFSET(STACK16FRAME,cs) */
 248         output( "\tmovl %%eax,0xbc(%%esp)\n" );  /* SegCs */
 249         output( "\tmovzwl 0x2c(%%edx),%%eax\n"); /* FIELD_OFFSET(STACK16FRAME,ip) */
 250         output( "\tmovl %%eax,0xb8(%%esp)\n" );  /* Eip */
 251
 252         output( "\tmovzwl %%fs:(%d), %%eax\n", STACKOFFSET+2 );
 253         output( "\tmovl %%eax,0xc8(%%esp)\n" );  /* SegSs */
 254         output( "\tmovzwl %%fs:(%d), %%eax\n", STACKOFFSET );
 255         output( "\taddl $0x2c,%%eax\n");         /* FIELD_OFFSET(STACK16FRAME,ip) */
 256         output( "\tmovl %%eax,0xc4(%%esp)\n" );  /* Esp */
 257 #if 0
 258         output( "\tfsave 0x1c(%%esp)\n" ); /* FloatSave */
 259 #endif
 260
 261         /* Push address of CONTEXT86 structure -- popped by the relay routine */
 262         output( "\tmovl %%esp,%%eax\n" );
 263         output( "\tandl $~15,%%esp\n" );
 264         output( "\tsubl $4,%%esp\n" );
 265         output( "\tpushl %%eax\n" );
 266     }
 267     else
 268     {
 269         output( "\tsubl $8,%%esp\n" );
 270         output( "\tandl $~15,%%esp\n" );
 271         output( "\taddl $8,%%esp\n" );
 272     }
 273
 274     /* Call relay routine (which will call the API entry point) */
 275     output( "\tleal 0x30(%%edx),%%eax\n" ); /* sizeof(STACK16FRAME) */
 276     output( "\tpushl %%eax\n" );
 277     output( "\tpushl 0x26(%%edx)\n");  /* FIELD_OFFSET(STACK16FRAME,entry_point) */
 278     output( "\tcall *0x20(%%edx)\n");  /* FIELD_OFFSET(STACK16FRAME,relay) */
 279
 280     if ( reg_func )
 281     {
 282         output( "\tleal -748(%%ebp),%%ebx\n" ); /* sizeof(CONTEXT) + FIELD_OFFSET(STACK32FRAME,ebp) */
 283
 284         /* Switch stack back */
 285         output( "\tmovw %%fs:(%d), %%ss\n", STACKOFFSET+2 );
 286         output( "\tmovzwl %%fs:(%d), %%esp\n", STACKOFFSET );
 287         output( "\tpopl %%fs:(%d)\n", STACKOFFSET );
 288
 289         /* Get return address to CallFrom16 stub */
 290         output( "\taddw $0x14,%%sp\n" ); /* FIELD_OFFSET(STACK16FRAME,callfrom_ip)-4 */
 291         output( "\tpopl %%eax\n" );
 292         output( "\tpopl %%edx\n" );
 293
 294         /* Restore all registers from CONTEXT */
 295         output( "\tmovw 0xc8(%%ebx),%%ss\n");   /* SegSs */
 296         output( "\tmovl 0xc4(%%ebx),%%esp\n");  /* Esp */
 297         output( "\taddl $4, %%esp\n" );  /* room for final return address */
 298
 299         output( "\tpushw 0xbc(%%ebx)\n");  /* SegCs */
 300         output( "\tpushw 0xb8(%%ebx)\n");  /* Eip */
 301         output( "\tpushl %%edx\n" );
 302         output( "\tpushl %%eax\n" );
 303         output( "\tpushl 0xc0(%%ebx)\n");  /* EFlags */
 304         output( "\tpushl 0x98(%%ebx)\n");  /* SegDs */
 305
 306         output( "\tpushl 0x94(%%ebx)\n");  /* SegEs */
 307         output( "\tpopl %%es\n" );
 308         output( "\tpushl 0x90(%%ebx)\n");  /* SegFs */
 309         output( "\tpopl %%fs\n" );
 310         output( "\tpushl 0x8c(%%ebx)\n");  /* SegGs */
 311         output( "\tpopl %%gs\n" );
 312
 313         output( "\tmovl 0xb4(%%ebx),%%ebp\n");  /* Ebp */
 314         output( "\tmovl 0xa0(%%ebx),%%esi\n");  /* Esi */
 315         output( "\tmovl 0x9c(%%ebx),%%edi\n");  /* Edi */
 316         output( "\tmovl 0xb0(%%ebx),%%eax\n");  /* Eax */
 317         output( "\tmovl 0xa8(%%ebx),%%edx\n");  /* Edx */
 318         output( "\tmovl 0xac(%%ebx),%%ecx\n");  /* Ecx */
 319         output( "\tmovl 0xa4(%%ebx),%%ebx\n");  /* Ebx */
 320
 321         output( "\tpopl %%ds\n" );
 322         output( "\tpopfl\n" );
 323         output( "\tlret\n" );
 324
 325         output_function_size( "__wine_call_from_16_regs" );
 326     }
 327     else
 328     {
 329         /* Switch stack back */
 330         output( "\tmovw %%fs:(%d), %%ss\n", STACKOFFSET+2 );
 331         output( "\tmovzwl %%fs:(%d), %%esp\n", STACKOFFSET );
 332         output( "\tpopl %%fs:(%d)\n", STACKOFFSET );
 333
 334         /* Restore registers */
 335         output( "\tpopl %%edx\n" );
 336         output( "\tpopl %%ecx\n" );
 337         output( "\tpopl %%ebp\n" );
 338         output( "\tpopw %%ds\n" );
 339         output( "\tpopw %%es\n" );
 340         output( "\tpopw %%fs\n" );
 341         output( "\tpopw %%gs\n" );
 342
 343         /* Return to return stub which will return to caller */
 344         output( "\tlret $12\n" );
 345
 346         output_function_size( "__wine_call_from_16" );
 347     }
 348 }
 349
 350
 351 /*******************************************************************
 352  *         BuildCallTo16Core
 353  *
 354  * This routine builds the core routines used in 32->16 thunks:
 355  *
 356  * extern DWORD WINAPI wine_call_to_16( FARPROC16 target, DWORD cbArgs, PEXCEPTION_HANDLER handler );
 357  * extern void WINAPI wine_call_to_16_regs( CONTEXT86 *context, DWORD cbArgs, PEXCEPTION_HANDLER handler );
 358  *
 359  * These routines can be called directly from 32-bit code.
 360  *
 361  * All routines expect that the 16-bit stack contents (arguments) and the
 362  * return address (segptr to CallTo16_Ret) were already set up by the
 363  * caller; nb_args must contain the number of bytes to be conserved.  The
 364  * 16-bit SS:SP will be set accordingly.
 365  *
 366  * All other registers are either taken from the CONTEXT86 structure
 367  * or else set to default values.  The target routine address is either
 368  * given directly or taken from the CONTEXT86.
 369  */
 370 static void BuildCallTo16Core( int reg_func )
 371 {
 372     const char *name = reg_func ? "wine_call_to_16_regs" : "wine_call_to_16";
 373     const char *func_name = is_pe() ? strmake( "%s@12", name ) : name;
 374
 375     /* Function header */
 376     output_function_header( func_name, 1 );
 377
 378     /* Function entry sequence */
 379     output_cfi( ".cfi_startproc" );
 380     output( "\tpushl %%ebp\n" );
 381     output_cfi( ".cfi_adjust_cfa_offset 4" );
 382     output_cfi( ".cfi_rel_offset %%ebp,0" );
 383     output( "\tmovl %%esp, %%ebp\n" );
 384     output_cfi( ".cfi_def_cfa_register %%ebp" );
 385
 386     /* Save the 32-bit registers */
 387     output( "\tpushl %%ebx\n" );
 388     output_cfi( ".cfi_rel_offset %%ebx,-4" );
 389     output( "\tpushl %%esi\n" );
 390     output_cfi( ".cfi_rel_offset %%esi,-8" );
 391     output( "\tpushl %%edi\n" );
 392     output_cfi( ".cfi_rel_offset %%edi,-12" );
 393     output( "\tmov %%gs,%%fs:(%d)\n", GS_OFFSET );
 394
 395     /* Setup exception frame */
 396     output( "\tpushl %%fs:(%d)\n", STACKOFFSET );
 397     output( "\tpushl 16(%%ebp)\n" ); /* handler */
 398     output( "\tpushl %%fs:(0)\n" );
 399     output( "\tmovl %%esp,%%fs:(0)\n" );
 400
 401     /* Call the actual CallTo16 routine (simulate a lcall) */
 402     output( "\tpushl %%cs\n" );
 403     output( "\tcall .L%s\n", name );
 404
 405     /* Remove exception frame */
 406     output( "\tpopl %%fs:(0)\n" );
 407     output( "\taddl $4, %%esp\n" );
 408     output( "\tpopl %%fs:(%d)\n", STACKOFFSET );
 409
 410     if ( !reg_func )
 411     {
 412         /* Convert return value */
 413         output( "\tandl $0xffff,%%eax\n" );
 414         output( "\tshll $16,%%edx\n" );
 415         output( "\torl %%edx,%%eax\n" );
 416     }
 417     else
 418     {
 419         /*
 420          * Modify CONTEXT86 structure to contain new values
 421          *
 422          * NOTE:  We restore only EAX, EBX, ECX, EDX, EBP, and ESP.
 423          *        The segment registers as well as ESI and EDI should
 424          *        not be modified by a well-behaved 16-bit routine in
 425          *        any case.  [If necessary, we could restore them as well,
 426          *        at the cost of a somewhat less efficient return path.]
 427          */
 428
 429         output( "\tmovl 0x14(%%esp),%%edi\n" ); /* FIELD_OFFSET(STACK32FRAME,target) - FIELD_OFFSET(STACK32FRAME,edi) */
 430                 /* everything above edi has been popped already */
 431
 432         output( "\tmovl %%eax,0xb0(%%edi)\n");  /* Eax */
 433         output( "\tmovl %%ebx,0xa4(%%edi)\n");  /* Ebx */
 434         output( "\tmovl %%ecx,0xac(%%edi)\n");  /* Ecx */
 435         output( "\tmovl %%edx,0xa8(%%edi)\n");  /* Edx */
 436         output( "\tmovl %%ebp,0xb4(%%edi)\n");  /* Ebp */
 437         output( "\tmovl %%esi,0xc4(%%edi)\n");  /* Esp */
 438                  /* The return glue code saved %esp into %esi */
 439     }
 440
 441     /* Restore the 32-bit registers */
 442     output( "\tpopl %%edi\n" );
 443     output_cfi( ".cfi_same_value %%edi" );
 444     output( "\tpopl %%esi\n" );
 445     output_cfi( ".cfi_same_value %%esi" );
 446     output( "\tpopl %%ebx\n" );
 447     output_cfi( ".cfi_same_value %%ebx" );
 448
 449     /* Function exit sequence */
 450     output( "\tpopl %%ebp\n" );
 451     output_cfi( ".cfi_def_cfa %%esp,4" );
 452     output_cfi( ".cfi_same_value %%ebp" );
 453     output( "\tret $12\n" );
 454     output_cfi( ".cfi_endproc" );
 455
 456
 457     /* Start of the actual CallTo16 routine */
 458
 459     output( ".L%s:\n", name );
 460
 461     /* Switch to the 16-bit stack */
 462     output( "\tmovl %%esp,%%edx\n" );
 463     output( "\tmovw %%fs:(%d),%%ss\n", STACKOFFSET + 2);
 464     output( "\tmovw %%fs:(%d),%%sp\n", STACKOFFSET );
 465     output( "\tmovl %%edx,%%fs:(%d)\n", STACKOFFSET );
 466
 467     /* Make %bp point to the previous stackframe (built by CallFrom16) */
 468     output( "\tmovzwl %%sp,%%ebp\n" );
 469     output( "\tleal 0x2a(%%ebp),%%ebp\n");  /* FIELD_OFFSET(STACK16FRAME,bp) */
 470
 471     /* Add the specified offset to the new sp */
 472     output( "\tsubw 0x2c(%%edx), %%sp\n");  /* FIELD_OFFSET(STACK32FRAME,nb_args) */
 473
 474     if (reg_func)
 475     {
 476         /* Push the called routine address */
 477         output( "\tmovl 0x28(%%edx),%%edx\n");  /* FIELD_OFFSET(STACK32FRAME,target) */
 478         output( "\tpushw 0xbc(%%edx)\n");  /* SegCs */
 479         output( "\tpushw 0xb8(%%edx)\n");  /* Eip */
 480
 481         /* Get the registers */
 482         output( "\tpushw 0x98(%%edx)\n");  /* SegDs */
 483         output( "\tpushl 0x94(%%edx)\n");  /* SegEs */
 484         output( "\tpopl %%es\n" );
 485         output( "\tmovl 0xb4(%%edx),%%ebp\n");  /* Ebp */
 486         output( "\tmovl 0xa0(%%edx),%%esi\n");  /* Esi */
 487         output( "\tmovl 0x9c(%%edx),%%edi\n");  /* Edi */
 488         output( "\tmovl 0xb0(%%edx),%%eax\n");  /* Eax */
 489         output( "\tmovl 0xa4(%%edx),%%ebx\n");  /* Ebx */
 490         output( "\tmovl 0xac(%%edx),%%ecx\n");  /* Ecx */
 491         output( "\tmovl 0xa8(%%edx),%%edx\n");  /* Edx */
 492
 493         /* Get the 16-bit ds */
 494         output( "\tpopw %%ds\n" );
 495     }
 496     else  /* not a register function */
 497     {
 498         /* Push the called routine address */
 499         output( "\tpushl 0x28(%%edx)\n"); /* FIELD_OFFSET(STACK32FRAME,target) */
 500
 501         /* Set %fs and %gs to the value saved by the last CallFrom16 */
 502         output( "\tpushw -22(%%ebp)\n" ); /* FIELD_OFFSET(STACK16FRAME,fs)-FIELD_OFFSET(STACK16FRAME,bp) */
 503         output( "\tpopw %%fs\n" );
 504         output( "\tpushw -20(%%ebp)\n" ); /* FIELD_OFFSET(STACK16FRAME,gs)-FIELD_OFFSET(STACK16FRAME,bp) */
 505         output( "\tpopw %%gs\n" );
 506
 507         /* Set %ds and %es (and %ax just in case) equal to %ss */
 508         output( "\tmovw %%ss,%%ax\n" );
 509         output( "\tmovw %%ax,%%ds\n" );
 510         output( "\tmovw %%ax,%%es\n" );
 511     }
 512
 513     /* Jump to the called routine */
 514     output( "\tlretw\n" );
 515
 516     /* Function footer */
 517     output_function_size( func_name );
 518 }
 519
 520
 521 /*******************************************************************
 522  *         BuildRet16Func
 523  *
 524  * Build the return code for 16-bit callbacks
 525  */
 526 static void BuildRet16Func(void)
 527 {
 528     output_function_header( "__wine_call_to_16_ret", 1 );
 529
 530     /* Save %esp into %esi */
 531     output( "\tmovl %%esp,%%esi\n" );
 532
 533     /* Restore 32-bit segment registers */
 534
 535     output( "\tmovl %%cs:%s", asm_name("CallTo16_DataSelector") );
 536     output( "-%s,%%edi\n", asm_name("__wine_call16_start") );
 537     output( "\tmovw %%di,%%ds\n" );
 538     output( "\tmovw %%di,%%es\n" );
 539
 540     output( "\tmov %%cs:%s", asm_name("CallTo16_TebSelector") );
 541     output( "-%s,%%fs\n", asm_name("__wine_call16_start") );
 542
 543     output( "\tmov %%fs:(%d),%%gs\n", GS_OFFSET );
 544
 545     /* Restore the 32-bit stack */
 546
 547     output( "\tmovw %%di,%%ss\n" );
 548     output( "\tmovl %%fs:(%d),%%esp\n", STACKOFFSET );
 549
 550     /* Return to caller */
 551
 552     output( "\tlret\n" );
 553     output_function_size( "__wine_call_to_16_ret" );
 554 }
 555
 556
 557 /*******************************************************************
 558  *         output_asm_relays16
 559  *
 560  * Build all the 16-bit relay callbacks
 561  */
 562 void output_asm_relays16(void)
 563 {
 564     /* File header */
 565
 566     output( "\t.text\n" );
 567     output( "%s:\n\n", asm_name("__wine_spec_thunk_text_16") );
 568
 569     output( "%s\n", asm_globl("__wine_call16_start") );
 570
 571     /* Standard CallFrom16 routine */
 572     BuildCallFrom16Core( 0, 0 );
 573
 574     /* Register CallFrom16 routine */
 575     BuildCallFrom16Core( 1, 0 );
 576
 577     /* C16ThkSL CallFrom16 routine */
 578     BuildCallFrom16Core( 0, 1 );
 579
 580     /* Standard CallTo16 routine */
 581     BuildCallTo16Core( 0 );
 582
 583     /* Register CallTo16 routine */
 584     BuildCallTo16Core( 1 );
 585
 586     /* Standard CallTo16 return stub */
 587     BuildRet16Func();
 588
 589     output( "%s\n", asm_globl("__wine_call16_end") );
 590     output_function_size( "__wine_spec_thunk_text_16" );
 591
 592     /* Declare the return address and data selector variables */
 593     output( "\n\t.data\n\t.balign 4\n" );
 594     output( "%s\n\t.long 0\n", asm_globl("CallTo16_DataSelector") );
 595     output( "%s\n\t.long 0\n", asm_globl("CallTo16_TebSelector") );
 596 }