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