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 #include <stdarg.h>
  30
  31 #include "build.h"
  32
  33 /* offset of the stack pointer relative to %fs:(0) */
  34 #define STACKOFFSET 0xc0  /* FIELD_OFFSET(TEB,WOW32Reserved) */
  35
  36 /* fix this if the ntdll_thread_regs structure is changed */
  37 #define GS_OFFSET  0x1d8  /* FIELD_OFFSET(TEB,SpareBytes1) + FIELD_OFFSET(ntdll_thread_data,gs) */
  38
  39 #define DPMI_VIF_OFFSET      (0x1fc + 0) /* FIELD_OFFSET(TEB,GdiTebBatch) + FIELD_OFFSET(WINE_VM86_TEB_INFO,dpmi_vif) */
  40 #define VM86_PENDING_OFFSET  (0x1fc + 4) /* FIELD_OFFSET(TEB,GdiTebBatch) + FIELD_OFFSET(WINE_VM86_TEB_INFO,vm86_pending) */
  41
  42 static void function_header( const char *name )
  43 {
  44     output( "\n\t.align %d\n", get_alignment(4) );
  45     output( "\t%s\n", func_declaration(name) );
  46     output( "%s\n", asm_globl(name) );
  47 }
  48
  49
  50 /*******************************************************************
  51  *         BuildCallFrom16Core
  52  *
  53  * This routine builds the core routines used in 16->32 thunks:
  54  * CallFrom16Word, CallFrom16Long, CallFrom16Register, and CallFrom16Thunk.
  55  *
  56  * These routines are intended to be called via a far call (with 32-bit
  57  * operand size) from 16-bit code.  The 16-bit code stub must push %bp,
  58  * the 32-bit entry point to be called, and the argument conversion
  59  * routine to be used (see stack layout below).
  60  *
  61  * The core routine completes the STACK16FRAME on the 16-bit stack and
  62  * switches to the 32-bit stack.  Then, the argument conversion routine
  63  * is called; it gets passed the 32-bit entry point and a pointer to the
  64  * 16-bit arguments (on the 16-bit stack) as parameters. (You can either
  65  * use conversion routines automatically generated by BuildCallFrom16,
  66  * or write your own for special purposes.)
  67  *
  68  * The conversion routine must call the 32-bit entry point, passing it
  69  * the converted arguments, and return its return value to the core.
  70  * After the conversion routine has returned, the core switches back
  71  * to the 16-bit stack, converts the return value to the DX:AX format
  72  * (CallFrom16Long), and returns to the 16-bit call stub.  All parameters,
  73  * including %bp, are popped off the stack.
  74  *
  75  * The 16-bit call stub now returns to the caller, popping the 16-bit
  76  * arguments if necessary (pascal calling convention).
  77  *
  78  * In the case of a 'register' function, CallFrom16Register fills a
  79  * CONTEXT86 structure with the values all registers had at the point
  80  * the first instruction of the 16-bit call stub was about to be
  81  * executed.  A pointer to this CONTEXT86 is passed as third parameter
  82  * to the argument conversion routine, which typically passes it on
  83  * to the called 32-bit entry point.
  84  *
  85  * CallFrom16Thunk is a special variant used by the implementation of
  86  * the Win95 16->32 thunk functions C16ThkSL and C16ThkSL01 and is
  87  * implemented as follows:
  88  * On entry, the EBX register is set up to contain a flat pointer to the
  89  * 16-bit stack such that EBX+22 points to the first argument.
  90  * Then, the entry point is called, while EBP is set up to point
  91  * to the return address (on the 32-bit stack).
  92  * The called function returns with CX set to the number of bytes
  93  * to be popped of the caller's stack.
  94  *
  95  * Stack layout upon entry to the core routine (STACK16FRAME):
  96  *  ...           ...
  97  * (sp+24) word   first 16-bit arg
  98  * (sp+22) word   cs
  99  * (sp+20) word   ip
 100  * (sp+18) word   bp
 101  * (sp+14) long   32-bit entry point (reused for Win16 mutex recursion count)
 102  * (sp+12) word   ip of actual entry point (necessary for relay debugging)
 103  * (sp+8)  long   relay (argument conversion) function entry point
 104  * (sp+4)  long   cs of 16-bit entry point
 105  * (sp)    long   ip of 16-bit entry point
 106  *
 107  * Added on the stack:
 108  * (sp-2)  word   saved gs
 109  * (sp-4)  word   saved fs
 110  * (sp-6)  word   saved es
 111  * (sp-8)  word   saved ds
 112  * (sp-12) long   saved ebp
 113  * (sp-16) long   saved ecx
 114  * (sp-20) long   saved edx
 115  * (sp-24) long   saved previous stack
 116  */
 117 static void BuildCallFrom16Core( int reg_func, int thunk )
 118 {
 119     /* Function header */
 120     if (thunk) function_header( "__wine_call_from_16_thunk" );
 121     else if (reg_func) function_header( "__wine_call_from_16_regs" );
 122     else function_header( "__wine_call_from_16" );
 123
 124     /* Create STACK16FRAME (except STACK32FRAME link) */
 125     output( "\tpushw %%gs\n" );
 126     output( "\tpushw %%fs\n" );
 127     output( "\tpushw %%es\n" );
 128     output( "\tpushw %%ds\n" );
 129     output( "\tpushl %%ebp\n" );
 130     output( "\tpushl %%ecx\n" );
 131     output( "\tpushl %%edx\n" );
 132
 133     /* Save original EFlags register */
 134     if (reg_func) output( "\tpushfl\n" );
 135
 136     if ( UsePIC )
 137     {
 138         output( "\tcall 1f\n" );
 139         output( "1:\tpopl %%ecx\n" );
 140         output( "\t.byte 0x2e\n\tmovl %s-1b(%%ecx),%%edx\n", asm_name("CallTo16_DataSelector") );
 141     }
 142     else
 143         output( "\t.byte 0x2e\n\tmovl %s,%%edx\n", asm_name("CallTo16_DataSelector") );
 144
 145     /* Load 32-bit segment registers */
 146     output( "\tmovw %%dx, %%ds\n" );
 147     output( "\tmovw %%dx, %%es\n" );
 148
 149     if ( UsePIC )
 150         output( "\tmovw %s-1b(%%ecx), %%fs\n", asm_name("CallTo16_TebSelector") );
 151     else
 152         output( "\tmovw %s, %%fs\n", asm_name("CallTo16_TebSelector") );
 153
 154     output( "\t.byte 0x64\n\tmov (%d),%%gs\n", GS_OFFSET );
 155
 156     /* Translate STACK16FRAME base to flat offset in %edx */
 157     output( "\tmovw %%ss, %%dx\n" );
 158     output( "\tandl $0xfff8, %%edx\n" );
 159     output( "\tshrl $1, %%edx\n" );
 160     if (UsePIC)
 161     {
 162         output( "\taddl wine_ldt_copy_ptr-1b(%%ecx),%%edx\n" );
 163         output( "\tmovl (%%edx), %%edx\n" );
 164     }
 165     else
 166         output( "\tmovl %s(%%edx), %%edx\n", asm_name("wine_ldt_copy") );
 167     output( "\tmovzwl %%sp, %%ebp\n" );
 168     output( "\tleal %d(%%ebp,%%edx), %%edx\n", reg_func ? 0 : -4 );
 169
 170     /* Get saved flags into %ecx */
 171     if (reg_func) output( "\tpopl %%ecx\n" );
 172
 173     /* Get the 32-bit stack pointer from the TEB and complete STACK16FRAME */
 174     output( "\t.byte 0x64\n\tmovl (%d), %%ebp\n", STACKOFFSET );
 175     output( "\tpushl %%ebp\n" );
 176
 177     /* Switch stacks */
 178     output( "\t.byte 0x64\n\tmovw %%ss, (%d)\n", STACKOFFSET + 2 );
 179     output( "\t.byte 0x64\n\tmovw %%sp, (%d)\n", STACKOFFSET );
 180     output( "\tpushl %%ds\n" );
 181     output( "\tpopl %%ss\n" );
 182     output( "\tmovl %%ebp, %%esp\n" );
 183     output( "\taddl $0x20,%%ebp\n");  /* FIELD_OFFSET(STACK32FRAME,ebp) */
 184
 185
 186     /* At this point:
 187        STACK16FRAME is completely set up
 188        DS, ES, SS: flat data segment
 189        FS: current TEB
 190        ESP: points to last STACK32FRAME
 191        EBP: points to ebp member of last STACK32FRAME
 192        EDX: points to current STACK16FRAME
 193        ECX: contains saved flags
 194        all other registers: unchanged */
 195
 196     /* Special case: C16ThkSL stub */
 197     if ( thunk )
 198     {
 199         /* Set up registers as expected and call thunk */
 200         output( "\tleal 0x1a(%%edx),%%ebx\n" );  /* sizeof(STACK16FRAME)-22 */
 201         output( "\tleal -4(%%esp), %%ebp\n" );
 202
 203         output( "\tcall *0x26(%%edx)\n");  /* FIELD_OFFSET(STACK16FRAME,entry_point) */
 204
 205         /* Switch stack back */
 206         output( "\t.byte 0x64\n\tmovw (%d), %%ss\n", STACKOFFSET+2 );
 207         output( "\t.byte 0x64\n\tmovzwl (%d), %%esp\n", STACKOFFSET );
 208         output( "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 209
 210         /* Restore registers and return directly to caller */
 211         output( "\taddl $8, %%esp\n" );
 212         output( "\tpopl %%ebp\n" );
 213         output( "\tpopw %%ds\n" );
 214         output( "\tpopw %%es\n" );
 215         output( "\tpopw %%fs\n" );
 216         output( "\tpopw %%gs\n" );
 217         output( "\taddl $20, %%esp\n" );
 218
 219         output( "\txorb %%ch, %%ch\n" );
 220         output( "\tpopl %%ebx\n" );
 221         output( "\taddw %%cx, %%sp\n" );
 222         output( "\tpush %%ebx\n" );
 223
 224         output( "\t.byte 0x66\n" );
 225         output( "\tlret\n" );
 226
 227         output_function_size( "__wine_call_from_16_thunk" );
 228         return;
 229     }
 230
 231
 232     /* Build register CONTEXT */
 233     if ( reg_func )
 234     {
 235         output( "\tsubl $0x2cc,%%esp\n" );       /* sizeof(CONTEXT86) */
 236
 237         output( "\tmovl %%ecx,0xc0(%%esp)\n" );  /* EFlags */
 238
 239         output( "\tmovl %%eax,0xb0(%%esp)\n" );  /* Eax */
 240         output( "\tmovl %%ebx,0xa4(%%esp)\n" );  /* Ebx */
 241         output( "\tmovl %%esi,0xa0(%%esp)\n" );  /* Esi */
 242         output( "\tmovl %%edi,0x9c(%%esp)\n" );  /* Edi */
 243
 244         output( "\tmovl 0x0c(%%edx),%%eax\n");   /* FIELD_OFFSET(STACK16FRAME,ebp) */
 245         output( "\tmovl %%eax,0xb4(%%esp)\n" );  /* Ebp */
 246         output( "\tmovl 0x08(%%edx),%%eax\n");   /* FIELD_OFFSET(STACK16FRAME,ecx) */
 247         output( "\tmovl %%eax,0xac(%%esp)\n" );  /* Ecx */
 248         output( "\tmovl 0x04(%%edx),%%eax\n");   /* FIELD_OFFSET(STACK16FRAME,edx) */
 249         output( "\tmovl %%eax,0xa8(%%esp)\n" );  /* Edx */
 250
 251         output( "\tmovzwl 0x10(%%edx),%%eax\n"); /* FIELD_OFFSET(STACK16FRAME,ds) */
 252         output( "\tmovl %%eax,0x98(%%esp)\n" );  /* SegDs */
 253         output( "\tmovzwl 0x12(%%edx),%%eax\n"); /* FIELD_OFFSET(STACK16FRAME,es) */
 254         output( "\tmovl %%eax,0x94(%%esp)\n" );  /* SegEs */
 255         output( "\tmovzwl 0x14(%%edx),%%eax\n"); /* FIELD_OFFSET(STACK16FRAME,fs) */
 256         output( "\tmovl %%eax,0x90(%%esp)\n" );  /* SegFs */
 257         output( "\tmovzwl 0x16(%%edx),%%eax\n"); /* FIELD_OFFSET(STACK16FRAME,gs) */
 258         output( "\tmovl %%eax,0x8c(%%esp)\n" );  /* SegGs */
 259
 260         output( "\tmovzwl 0x2e(%%edx),%%eax\n"); /* FIELD_OFFSET(STACK16FRAME,cs) */
 261         output( "\tmovl %%eax,0xbc(%%esp)\n" );  /* SegCs */
 262         output( "\tmovzwl 0x2c(%%edx),%%eax\n"); /* FIELD_OFFSET(STACK16FRAME,ip) */
 263         output( "\tmovl %%eax,0xb8(%%esp)\n" );  /* Eip */
 264
 265         output( "\t.byte 0x64\n\tmovzwl (%d), %%eax\n", STACKOFFSET+2 );
 266         output( "\tmovl %%eax,0xc8(%%esp)\n" );  /* SegSs */
 267         output( "\t.byte 0x64\n\tmovzwl (%d), %%eax\n", STACKOFFSET );
 268         output( "\taddl $0x2c,%%eax\n");         /* FIELD_OFFSET(STACK16FRAME,ip) */
 269         output( "\tmovl %%eax,0xc4(%%esp)\n" );  /* Esp */
 270 #if 0
 271         output( "\tfsave 0x1c(%%esp)\n" ); /* FloatSave */
 272 #endif
 273
 274         /* Push address of CONTEXT86 structure -- popped by the relay routine */
 275         output( "\tmovl %%esp,%%eax\n" );
 276         output( "\tandl $~15,%%esp\n" );
 277         output( "\tsubl $4,%%esp\n" );
 278         output( "\tpushl %%eax\n" );
 279     }
 280     else
 281     {
 282         output( "\tsubl $8,%%esp\n" );
 283         output( "\tandl $~15,%%esp\n" );
 284         output( "\taddl $8,%%esp\n" );
 285     }
 286
 287     /* Call relay routine (which will call the API entry point) */
 288     output( "\tleal 0x30(%%edx),%%eax\n" ); /* sizeof(STACK16FRAME) */
 289     output( "\tpushl %%eax\n" );
 290     output( "\tpushl 0x26(%%edx)\n");  /* FIELD_OFFSET(STACK16FRAME,entry_point) */
 291     output( "\tcall *0x20(%%edx)\n");  /* FIELD_OFFSET(STACK16FRAME,relay) */
 292
 293     if ( reg_func )
 294     {
 295         output( "\tleal -748(%%ebp),%%ebx\n" ); /* sizeof(CONTEXT) + FIELD_OFFSET(STACK32FRAME,ebp) */
 296
 297         /* Switch stack back */
 298         output( "\t.byte 0x64\n\tmovw (%d), %%ss\n", STACKOFFSET+2 );
 299         output( "\t.byte 0x64\n\tmovzwl (%d), %%esp\n", STACKOFFSET );
 300         output( "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 301
 302         /* Get return address to CallFrom16 stub */
 303         output( "\taddw $0x14,%%sp\n" ); /* FIELD_OFFSET(STACK16FRAME,callfrom_ip)-4 */
 304         output( "\tpopl %%eax\n" );
 305         output( "\tpopl %%edx\n" );
 306
 307         /* Restore all registers from CONTEXT */
 308         output( "\tmovw 0xc8(%%ebx),%%ss\n");   /* SegSs */
 309         output( "\tmovl 0xc4(%%ebx),%%esp\n");  /* Esp */
 310         output( "\taddl $4, %%esp\n" );  /* room for final return address */
 311
 312         output( "\tpushw 0xbc(%%ebx)\n");  /* SegCs */
 313         output( "\tpushw 0xb8(%%ebx)\n");  /* Eip */
 314         output( "\tpushl %%edx\n" );
 315         output( "\tpushl %%eax\n" );
 316         output( "\tpushl 0xc0(%%ebx)\n");  /* EFlags */
 317         output( "\tpushl 0x98(%%ebx)\n");  /* SegDs */
 318
 319         output( "\tpushl 0x94(%%ebx)\n");  /* SegEs */
 320         output( "\tpopl %%es\n" );
 321         output( "\tpushl 0x90(%%ebx)\n");  /* SegFs */
 322         output( "\tpopl %%fs\n" );
 323         output( "\tpushl 0x8c(%%ebx)\n");  /* SegGs */
 324         output( "\tpopl %%gs\n" );
 325
 326         output( "\tmovl 0xb4(%%ebx),%%ebp\n");  /* Ebp */
 327         output( "\tmovl 0xa0(%%ebx),%%esi\n");  /* Esi */
 328         output( "\tmovl 0x9c(%%ebx),%%edi\n");  /* Edi */
 329         output( "\tmovl 0xb0(%%ebx),%%eax\n");  /* Eax */
 330         output( "\tmovl 0xa8(%%ebx),%%edx\n");  /* Edx */
 331         output( "\tmovl 0xac(%%ebx),%%ecx\n");  /* Ecx */
 332         output( "\tmovl 0xa4(%%ebx),%%ebx\n");  /* Ebx */
 333
 334         output( "\tpopl %%ds\n" );
 335         output( "\tpopfl\n" );
 336         output( "\tlret\n" );
 337
 338         output_function_size( "__wine_call_from_16_regs" );
 339     }
 340     else
 341     {
 342         /* Switch stack back */
 343         output( "\t.byte 0x64\n\tmovw (%d), %%ss\n", STACKOFFSET+2 );
 344         output( "\t.byte 0x64\n\tmovzwl (%d), %%esp\n", STACKOFFSET );
 345         output( "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 346
 347         /* Restore registers */
 348         output( "\tpopl %%edx\n" );
 349         output( "\tpopl %%ecx\n" );
 350         output( "\tpopl %%ebp\n" );
 351         output( "\tpopw %%ds\n" );
 352         output( "\tpopw %%es\n" );
 353         output( "\tpopw %%fs\n" );
 354         output( "\tpopw %%gs\n" );
 355
 356         /* Return to return stub which will return to caller */
 357         output( "\tlret $12\n" );
 358
 359         output_function_size( "__wine_call_from_16" );
 360     }
 361 }
 362
 363
 364 /*******************************************************************
 365  *         BuildCallTo16Core
 366  *
 367  * This routine builds the core routines used in 32->16 thunks:
 368  *
 369  * extern DWORD WINAPI wine_call_to_16( FARPROC16 target, DWORD cbArgs, PEXCEPTION_HANDLER handler );
 370  * extern void WINAPI wine_call_to_16_regs( CONTEXT86 *context, DWORD cbArgs, PEXCEPTION_HANDLER handler );
 371  *
 372  * These routines can be called directly from 32-bit code.
 373  *
 374  * All routines expect that the 16-bit stack contents (arguments) and the
 375  * return address (segptr to CallTo16_Ret) were already set up by the
 376  * caller; nb_args must contain the number of bytes to be conserved.  The
 377  * 16-bit SS:SP will be set accordingly.
 378  *
 379  * All other registers are either taken from the CONTEXT86 structure
 380  * or else set to default values.  The target routine address is either
 381  * given directly or taken from the CONTEXT86.
 382  */
 383 static void BuildCallTo16Core( int reg_func )
 384 {
 385     const char *name = reg_func ? "wine_call_to_16_regs" : "wine_call_to_16";
 386
 387     /* Function header */
 388     function_header( name );
 389
 390     /* Function entry sequence */
 391     output_cfi( ".cfi_startproc" );
 392     output( "\tpushl %%ebp\n" );
 393     output_cfi( ".cfi_adjust_cfa_offset 4" );
 394     output_cfi( ".cfi_rel_offset %%ebp,0" );
 395     output( "\tmovl %%esp, %%ebp\n" );
 396     output_cfi( ".cfi_def_cfa_register %%ebp" );
 397
 398     /* Save the 32-bit registers */
 399     output( "\tpushl %%ebx\n" );
 400     output_cfi( ".cfi_rel_offset %%ebx,-4" );
 401     output( "\tpushl %%esi\n" );
 402     output_cfi( ".cfi_rel_offset %%esi,-8" );
 403     output( "\tpushl %%edi\n" );
 404     output_cfi( ".cfi_rel_offset %%edi,-12" );
 405     output( "\t.byte 0x64\n\tmov %%gs,(%d)\n", GS_OFFSET );
 406
 407     /* Setup exception frame */
 408     output( "\t.byte 0x64\n\tpushl (%d)\n", STACKOFFSET );
 409     output( "\tpushl 16(%%ebp)\n" ); /* handler */
 410     output( "\t.byte 0x64\n\tpushl (0)\n" );
 411     output( "\t.byte 0x64\n\tmovl %%esp,(0)\n" );
 412
 413     /* Call the actual CallTo16 routine (simulate a lcall) */
 414     output( "\tpushl %%cs\n" );
 415     output( "\tcall .L%s\n", name );
 416
 417     /* Remove exception frame */
 418     output( "\t.byte 0x64\n\tpopl (0)\n" );
 419     output( "\taddl $4, %%esp\n" );
 420     output( "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 421
 422     if ( !reg_func )
 423     {
 424         /* Convert return value */
 425         output( "\tandl $0xffff,%%eax\n" );
 426         output( "\tshll $16,%%edx\n" );
 427         output( "\torl %%edx,%%eax\n" );
 428     }
 429     else
 430     {
 431         /*
 432          * Modify CONTEXT86 structure to contain new values
 433          *
 434          * NOTE:  We restore only EAX, EBX, EDX, EDX, EBP, and ESP.
 435          *        The segment registers as well as ESI and EDI should
 436          *        not be modified by a well-behaved 16-bit routine in
 437          *        any case.  [If necessary, we could restore them as well,
 438          *        at the cost of a somewhat less efficient return path.]
 439          */
 440
 441         output( "\tmovl 0x14(%%esp),%%edi\n" ); /* FIELD_OFFSET(STACK32FRAME,target) - FIELD_OFFSET(STACK32FRAME,edi) */
 442                 /* everything above edi has been popped already */
 443
 444         output( "\tmovl %%eax,0xb0(%%edi)\n");  /* Eax */
 445         output( "\tmovl %%ebx,0xa4(%%edi)\n");  /* Ebx */
 446         output( "\tmovl %%ecx,0xac(%%edi)\n");  /* Ecx */
 447         output( "\tmovl %%edx,0xa8(%%edi)\n");  /* Edx */
 448         output( "\tmovl %%ebp,0xb4(%%edi)\n");  /* Ebp */
 449         output( "\tmovl %%esi,0xc4(%%edi)\n");  /* Esp */
 450                  /* The return glue code saved %esp into %esi */
 451     }
 452
 453     /* Restore the 32-bit registers */
 454     output( "\tpopl %%edi\n" );
 455     output_cfi( ".cfi_same_value %%edi" );
 456     output( "\tpopl %%esi\n" );
 457     output_cfi( ".cfi_same_value %%esi" );
 458     output( "\tpopl %%ebx\n" );
 459     output_cfi( ".cfi_same_value %%ebx" );
 460
 461     /* Function exit sequence */
 462     output( "\tpopl %%ebp\n" );
 463     output_cfi( ".cfi_def_cfa %%esp,4" );
 464     output_cfi( ".cfi_same_value %%ebp" );
 465     output( "\tret $12\n" );
 466     output_cfi( ".cfi_endproc" );
 467
 468
 469     /* Start of the actual CallTo16 routine */
 470
 471     output( ".L%s:\n", name );
 472
 473     /* Switch to the 16-bit stack */
 474     output( "\tmovl %%esp,%%edx\n" );
 475     output( "\t.byte 0x64\n\tmovw (%d),%%ss\n", STACKOFFSET + 2);
 476     output( "\t.byte 0x64\n\tmovw (%d),%%sp\n", STACKOFFSET );
 477     output( "\t.byte 0x64\n\tmovl %%edx,(%d)\n", STACKOFFSET );
 478
 479     /* Make %bp point to the previous stackframe (built by CallFrom16) */
 480     output( "\tmovzwl %%sp,%%ebp\n" );
 481     output( "\tleal 0x2a(%%ebp),%%ebp\n");  /* FIELD_OFFSET(STACK16FRAME,bp) */
 482
 483     /* Add the specified offset to the new sp */
 484     output( "\tsubw 0x2c(%%edx), %%sp\n");  /* FIELD_OFFSET(STACK32FRAME,nb_args) */
 485
 486     if (reg_func)
 487     {
 488         /* Push the called routine address */
 489         output( "\tmovl 0x28(%%edx),%%edx\n");  /* FIELD_OFFSET(STACK32FRAME,target) */
 490         output( "\tpushw 0xbc(%%edx)\n");  /* SegCs */
 491         output( "\tpushw 0xb8(%%edx)\n");  /* Eip */
 492
 493         /* Get the registers */
 494         output( "\tpushw 0x98(%%edx)\n");  /* SegDs */
 495         output( "\tpushl 0x94(%%edx)\n");  /* SegEs */
 496         output( "\tpopl %%es\n" );
 497         output( "\tpushl 0x90(%%edx)\n");  /* SegFs */
 498         output( "\tpopl %%fs\n" );
 499         output( "\tpushl 0x8c(%%edx)\n");  /* SegGs */
 500         output( "\tpopl %%gs\n" );
 501         output( "\tmovl 0xb4(%%edx),%%ebp\n");  /* Ebp */
 502         output( "\tmovl 0xa0(%%edx),%%esi\n");  /* Esi */
 503         output( "\tmovl 0x9c(%%edx),%%edi\n");  /* Edi */
 504         output( "\tmovl 0xb0(%%edx),%%eax\n");  /* Eax */
 505         output( "\tmovl 0xa4(%%edx),%%ebx\n");  /* Ebx */
 506         output( "\tmovl 0xac(%%edx),%%ecx\n");  /* Ecx */
 507         output( "\tmovl 0xa8(%%edx),%%edx\n");  /* Edx */
 508
 509         /* Get the 16-bit ds */
 510         output( "\tpopw %%ds\n" );
 511     }
 512     else  /* not a register function */
 513     {
 514         /* Push the called routine address */
 515         output( "\tpushl 0x28(%%edx)\n"); /* FIELD_OFFSET(STACK32FRAME,target) */
 516
 517         /* Set %fs and %gs to the value saved by the last CallFrom16 */
 518         output( "\tpushw -22(%%ebp)\n" ); /* FIELD_OFFSET(STACK16FRAME,fs)-FIELD_OFFSET(STACK16FRAME,bp) */
 519         output( "\tpopw %%fs\n" );
 520         output( "\tpushw -20(%%ebp)\n" ); /* FIELD_OFFSET(STACK16FRAME,gs)-FIELD_OFFSET(STACK16FRAME,bp) */
 521         output( "\tpopw %%gs\n" );
 522
 523         /* Set %ds and %es (and %ax just in case) equal to %ss */
 524         output( "\tmovw %%ss,%%ax\n" );
 525         output( "\tmovw %%ax,%%ds\n" );
 526         output( "\tmovw %%ax,%%es\n" );
 527     }
 528
 529     /* Jump to the called routine */
 530     output( "\t.byte 0x66\n" );
 531     output( "\tlret\n" );
 532
 533     /* Function footer */
 534     output_function_size( name );
 535 }
 536
 537
 538 /*******************************************************************
 539  *         BuildRet16Func
 540  *
 541  * Build the return code for 16-bit callbacks
 542  */
 543 static void BuildRet16Func(void)
 544 {
 545     function_header( "__wine_call_to_16_ret" );
 546
 547     /* Save %esp into %esi */
 548     output( "\tmovl %%esp,%%esi\n" );
 549
 550     /* Restore 32-bit segment registers */
 551
 552     output( "\t.byte 0x2e\n\tmovl %s", asm_name("CallTo16_DataSelector") );
 553     output( "-%s,%%edi\n", asm_name("__wine_call16_start") );
 554     output( "\tmovw %%di,%%ds\n" );
 555     output( "\tmovw %%di,%%es\n" );
 556
 557     output( "\t.byte 0x2e\n\tmov %s", asm_name("CallTo16_TebSelector") );
 558     output( "-%s,%%fs\n", asm_name("__wine_call16_start") );
 559
 560     output( "\t.byte 0x64\n\tmov (%d),%%gs\n", GS_OFFSET );
 561
 562     /* Restore the 32-bit stack */
 563
 564     output( "\tmovw %%di,%%ss\n" );
 565     output( "\t.byte 0x64\n\tmovl (%d),%%esp\n", STACKOFFSET );
 566
 567     /* Return to caller */
 568
 569     output( "\tlret\n" );
 570     output_function_size( "__wine_call_to_16_ret" );
 571 }
 572
 573
 574 /*******************************************************************
 575  *         BuildCallTo32CBClient
 576  *
 577  * Call a CBClient relay stub from 32-bit code (KERNEL.620).
 578  *
 579  * Since the relay stub is itself 32-bit, this should not be a problem;
 580  * unfortunately, the relay stubs are expected to switch back to a
 581  * 16-bit stack (and 16-bit code) after completion :-(
 582  *
 583  * This would conflict with our 16- vs. 32-bit stack handling, so
 584  * we simply switch *back* to our 32-bit stack before returning to
 585  * the caller ...
 586  *
 587  * The CBClient relay stub expects to be called with the following
 588  * 16-bit stack layout, and with ebp and ebx pointing into the 16-bit
 589  * stack at the designated places:
 590  *
 591  *    ...
 592  *  (ebp+14) original arguments to the callback routine
 593  *  (ebp+10) far return address to original caller
 594  *  (ebp+6)  Thunklet target address
 595  *  (ebp+2)  Thunklet relay ID code
 596  *  (ebp)    BP (saved by CBClientGlueSL)
 597  *  (ebp-2)  SI (saved by CBClientGlueSL)
 598  *  (ebp-4)  DI (saved by CBClientGlueSL)
 599  *  (ebp-6)  DS (saved by CBClientGlueSL)
 600  *
 601  *   ...     buffer space used by the 16-bit side glue for temp copies
 602  *
 603  *  (ebx+4)  far return address to 16-bit side glue code
 604  *  (ebx)    saved 16-bit ss:sp (pointing to ebx+4)
 605  *
 606  * The 32-bit side glue code accesses both the original arguments (via ebp)
 607  * and the temporary copies prepared by the 16-bit side glue (via ebx).
 608  * After completion, the stub will load ss:sp from the buffer at ebx
 609  * and perform a far return to 16-bit code.
 610  *
 611  * To trick the relay stub into returning to us, we replace the 16-bit
 612  * return address to the glue code by a cs:ip pair pointing to our
 613  * return entry point (the original return address is saved first).
 614  * Our return stub thus called will then reload the 32-bit ss:esp and
 615  * return to 32-bit code (by using and ss:esp value that we have also
 616  * pushed onto the 16-bit stack before and a cs:eip values found at
 617  * that position on the 32-bit stack).  The ss:esp to be restored is
 618  * found relative to the 16-bit stack pointer at:
 619  *
 620  *  (ebx-4)   ss  (flat)
 621  *  (ebx-8)   sp  (32-bit stack pointer)
 622  *
 623  * The second variant of this routine, CALL32_CBClientEx, which is used
 624  * to implement KERNEL.621, has to cope with yet another problem: Here,
 625  * the 32-bit side directly returns to the caller of the CBClient thunklet,
 626  * restoring registers saved by CBClientGlueSL and cleaning up the stack.
 627  * As we have to return to our 32-bit code first, we have to adapt the
 628  * layout of our temporary area so as to include values for the registers
 629  * that are to be restored, and later (in the implementation of KERNEL.621)
 630  * we *really* restore them. The return stub restores DS, DI, SI, and BP
 631  * from the stack, skips the next 8 bytes (CBClient relay code / target),
 632  * and then performs a lret NN, where NN is the number of arguments to be
 633  * removed. Thus, we prepare our temporary area as follows:
 634  *
 635  *     (ebx+22) 16-bit cs  (this segment)
 636  *     (ebx+20) 16-bit ip  ('16-bit' return entry point)
 637  *     (ebx+16) 32-bit ss  (flat)
 638  *     (ebx+12) 32-bit sp  (32-bit stack pointer)
 639  *     (ebx+10) 16-bit bp  (points to ebx+24)
 640  *     (ebx+8)  16-bit si  (ignored)
 641  *     (ebx+6)  16-bit di  (ignored)
 642  *     (ebx+4)  16-bit ds  (we actually use the flat DS here)
 643  *     (ebx+2)  16-bit ss  (16-bit stack segment)
 644  *     (ebx+0)  16-bit sp  (points to ebx+4)
 645  *
 646  * Note that we ensure that DS is not changed and remains the flat segment,
 647  * and the 32-bit stack pointer our own return stub needs fits just
 648  * perfectly into the 8 bytes that are skipped by the Windows stub.
 649  * One problem is that we have to determine the number of removed arguments,
 650  * as these have to be really removed in KERNEL.621. Thus, the BP value
 651  * that we place in the temporary area to be restored, contains the value
 652  * that SP would have if no arguments were removed. By comparing the actual
 653  * value of SP with this value in our return stub we can compute the number
 654  * of removed arguments. This is then returned to KERNEL.621.
 655  *
 656  * The stack layout of this function:
 657  * (ebp+20)  nArgs     pointer to variable receiving nr. of args (Ex only)
 658  * (ebp+16)  esi       pointer to caller's esi value
 659  * (ebp+12)  arg       ebp value to be set for relay stub
 660  * (ebp+8)   func      CBClient relay stub address
 661  * (ebp+4)   ret addr
 662  * (ebp)     ebp
 663  */
 664 static void BuildCallTo32CBClient( int isEx )
 665 {
 666     function_header( isEx ? "CALL32_CBClientEx" : "CALL32_CBClient" );
 667
 668     /* Entry code */
 669
 670     output_cfi( ".cfi_startproc" );
 671     output( "\tpushl %%ebp\n" );
 672     output_cfi( ".cfi_adjust_cfa_offset 4" );
 673     output_cfi( ".cfi_rel_offset %%ebp,0" );
 674     output( "\tmovl %%esp,%%ebp\n" );
 675     output_cfi( ".cfi_def_cfa_register %%ebp" );
 676     output( "\tpushl %%edi\n" );
 677     output_cfi( ".cfi_rel_offset %%edi,-4" );
 678     output( "\tpushl %%esi\n" );
 679     output_cfi( ".cfi_rel_offset %%esi,-8" );
 680     output( "\tpushl %%ebx\n" );
 681     output_cfi( ".cfi_rel_offset %%ebx,-12" );
 682
 683     /* Get pointer to temporary area and save the 32-bit stack pointer */
 684
 685     output( "\tmovl 16(%%ebp), %%ebx\n" );
 686     output( "\tleal -8(%%esp), %%eax\n" );
 687
 688     if ( !isEx )
 689         output( "\tmovl %%eax, -8(%%ebx)\n" );
 690     else
 691         output( "\tmovl %%eax, 12(%%ebx)\n" );
 692
 693     /* Set up registers and call CBClient relay stub (simulating a far call) */
 694
 695     output( "\tmovl 20(%%ebp), %%esi\n" );
 696     output( "\tmovl (%%esi), %%esi\n" );
 697
 698     output( "\tmovl 8(%%ebp), %%eax\n" );
 699     output( "\tmovl 12(%%ebp), %%ebp\n" );
 700
 701     output( "\tpushl %%cs\n" );
 702     output( "\tcall *%%eax\n" );
 703
 704     /* Return new esi value to caller */
 705
 706     output( "\tmovl 32(%%esp), %%edi\n" );
 707     output( "\tmovl %%esi, (%%edi)\n" );
 708
 709     /* Return argument size to caller */
 710     if ( isEx )
 711     {
 712         output( "\tmovl 36(%%esp), %%ebx\n" );
 713         output( "\tmovl %%ebp, (%%ebx)\n" );
 714     }
 715
 716     /* Restore registers and return */
 717
 718     output( "\tpopl %%ebx\n" );
 719     output_cfi( ".cfi_same_value %%ebx" );
 720     output( "\tpopl %%esi\n" );
 721     output_cfi( ".cfi_same_value %%esi" );
 722     output( "\tpopl %%edi\n" );
 723     output_cfi( ".cfi_same_value %%edi" );
 724     output( "\tpopl %%ebp\n" );
 725     output_cfi( ".cfi_def_cfa %%esp,4" );
 726     output_cfi( ".cfi_same_value %%ebp" );
 727     output( "\tret\n" );
 728     output_cfi( ".cfi_endproc" );
 729     output_function_size( isEx ? "CALL32_CBClientEx" : "CALL32_CBClient" );
 730
 731     /* '16-bit' return stub */
 732
 733     function_header( isEx ? "CALL32_CBClientEx_Ret" : "CALL32_CBClient_Ret" );
 734     if ( !isEx )
 735     {
 736         output( "\tmovzwl %%sp, %%ebx\n" );
 737         output( "\tlssl %%ss:-16(%%ebx), %%esp\n" );
 738     }
 739     else
 740     {
 741         output( "\tmovzwl %%bp, %%ebx\n" );
 742         output( "\tsubw %%bp, %%sp\n" );
 743         output( "\tmovzwl %%sp, %%ebp\n" );
 744         output( "\tlssl %%ss:-12(%%ebx), %%esp\n" );
 745     }
 746     output( "\tlret\n" );
 747     output_function_size( isEx ? "CALL32_CBClientEx_Ret" : "CALL32_CBClient_Ret" );
 748 }
 749
 750
 751 /*******************************************************************
 752  *         build_call_from_regs_x86
 753  *
 754  * Build a 32-bit-to-Wine call-back function for a 'register' function.
 755  * 'args' is the number of dword arguments.
 756  *
 757  * Stack layout:
 758  *   ...
 759  * (ebp+20)  first arg
 760  * (ebp+16)  ret addr to user code
 761  * (ebp+12)  func to call (relative to relay code ret addr)
 762  * (ebp+8)   number of args
 763  * (ebp+4)   ret addr to relay code
 764  * (ebp+0)   saved ebp
 765  * (ebp-128) buffer area to allow stack frame manipulation
 766  * (ebp-332) CONTEXT86 struct
 767  * (ebp-336) padding for stack alignment
 768  * (ebp-336-n) CONTEXT86 *argument
 769  *  ....     other arguments copied from (ebp+12)
 770  *
 771  * The entry point routine is called with a CONTEXT* extra argument,
 772  * following the normal args. In this context structure, EIP_reg
 773  * contains the return address to user code, and ESP_reg the stack
 774  * pointer on return (with the return address and arguments already
 775  * removed).
 776  */
 777 static void build_call_from_regs_x86(void)
 778 {
 779     static const int STACK_SPACE = 128 + 0x2cc /* sizeof(CONTEXT86) */;
 780
 781     /* Function header */
 782
 783     output( "\t.text\n" );
 784     function_header( "__wine_call_from_regs" );
 785
 786     /* Allocate some buffer space on the stack */
 787
 788     output_cfi( ".cfi_startproc" );
 789     output( "\tpushl %%ebp\n" );
 790     output_cfi( ".cfi_adjust_cfa_offset 4" );
 791     output_cfi( ".cfi_rel_offset %%ebp,0" );
 792     output( "\tmovl %%esp,%%ebp\n" );
 793     output_cfi( ".cfi_def_cfa_register %%ebp" );
 794     output( "\tleal -%d(%%esp),%%esp\n", STACK_SPACE );
 795
 796     /* Build the context structure */
 797
 798     output( "\tmovl %%eax,0xb0(%%esp)\n" );  /* Eax */
 799     output( "\tpushfl\n" );
 800     output( "\tpopl %%eax\n" );
 801     output( "\tmovl %%eax,0xc0(%%esp)\n");  /* EFlags */
 802     output( "\tmovl 0(%%ebp),%%eax\n" );
 803     output( "\tmovl %%eax,0xb4(%%esp)\n");  /* Ebp */
 804     output( "\tmovl %%ebx,0xa4(%%esp)\n");  /* Ebx */
 805     output( "\tmovl %%ecx,0xac(%%esp)\n");  /* Ecx */
 806     output( "\tmovl %%edx,0xa8(%%esp)\n");  /* Edx */
 807     output( "\tmovl %%esi,0xa0(%%esp)\n");  /* Esi */
 808     output( "\tmovl %%edi,0x9c(%%esp)\n");  /* Edi */
 809
 810     output( "\txorl %%eax,%%eax\n" );
 811     output( "\tmovw %%cs,%%ax\n" );
 812     output( "\tmovl %%eax,0xbc(%%esp)\n");  /* SegCs */
 813     output( "\tmovw %%es,%%ax\n" );
 814     output( "\tmovl %%eax,0x94(%%esp)\n");  /* SegEs */
 815     output( "\tmovw %%fs,%%ax\n" );
 816     output( "\tmovl %%eax,0x90(%%esp)\n");  /* SegFs */
 817     output( "\tmovw %%gs,%%ax\n" );
 818     output( "\tmovl %%eax,0x8c(%%esp)\n");  /* SegGs */
 819     output( "\tmovw %%ss,%%ax\n" );
 820     output( "\tmovl %%eax,0xc8(%%esp)\n");  /* SegSs */
 821     output( "\tmovw %%ds,%%ax\n" );
 822     output( "\tmovl %%eax,0x98(%%esp)\n");  /* SegDs */
 823     output( "\tmovw %%ax,%%es\n" );  /* set %es equal to %ds just in case */
 824
 825     output( "\tmovl $0x10007,0(%%esp)\n");  /* ContextFlags */
 826
 827     output( "\tmovl 16(%%ebp),%%eax\n" ); /* Get %eip at time of call */
 828     output( "\tmovl %%eax,0xb8(%%esp)\n");  /* Eip */
 829
 830     /* Transfer the arguments */
 831
 832     output( "\tmovl 8(%%ebp),%%ecx\n" );    /* fetch number of args to copy */
 833     output( "\tleal 4(,%%ecx,4),%%edx\n" ); /* add 4 for context arg */
 834     output( "\tsubl %%edx,%%esp\n" );
 835     output( "\tandl $~15,%%esp\n" );
 836     output( "\tleal 20(%%ebp),%%esi\n" );  /* get %esp at time of call */
 837     output( "\tmovl %%esp,%%edi\n" );
 838     output( "\ttest %%ecx,%%ecx\n" );
 839     output( "\tjz 1f\n" );
 840     output( "\tcld\n" );
 841     output( "\trep\n\tmovsl\n" );  /* copy args */
 842     output( "1:\tleal %d(%%ebp),%%eax\n", -STACK_SPACE );  /* get addr of context struct */
 843     output( "\tmovl %%eax,(%%edi)\n" );    /* and pass it as extra arg */
 844     output( "\tmovl %%esi,%d(%%ebp)\n", 0xc4 /* Esp */ - STACK_SPACE );
 845
 846     /* Call the entry point */
 847
 848     output( "\tmovl 4(%%ebp),%%eax\n" );   /* get relay code addr */
 849     output( "\taddl 12(%%ebp),%%eax\n" );
 850     output( "\tcall *%%eax\n" );
 851     output( "\tleal -%d(%%ebp),%%ecx\n", STACK_SPACE );
 852
 853     /* Restore the context structure */
 854
 855     output( "2:\tpushl 0x94(%%ecx)\n" );    /* SegEs */
 856     output( "\tpopl %%es\n" );
 857     output( "\tpushl 0x90(%%ecx)\n" );      /* SegFs */
 858     output( "\tpopl %%fs\n" );
 859     output( "\tpushl 0x8c(%%ecx)\n" );      /* SegGs */
 860     output( "\tpopl %%gs\n" );
 861
 862     output( "\tmovw %%ss,%%ax\n" );
 863     output( "\tcmpw 0xc8(%%ecx),%%ax\n" );  /* SegSs */
 864     output( "\tjne 3f\n" );
 865
 866     /* As soon as we have switched stacks the context structure could
 867      * be invalid (when signal handlers are executed for example). Copy
 868      * values on the target stack before changing ESP. */
 869
 870     output( "\tmovl 0xc4(%%ecx),%%eax\n" ); /* Esp */
 871     output( "\tleal -4*4(%%eax),%%eax\n" );
 872
 873     output( "\tmovl 0xc0(%%ecx),%%edx\n" ); /* EFlags */
 874     output( "\t.byte 0x36\n\tmovl %%edx,3*4(%%eax)\n" );
 875     output( "\tmovl 0xbc(%%ecx),%%edx\n" ); /* SegCs */
 876     output( "\t.byte 0x36\n\tmovl %%edx,2*4(%%eax)\n" );
 877     output( "\tmovl 0xb8(%%ecx),%%edx\n" ); /* Eip */
 878     output( "\t.byte 0x36\n\tmovl %%edx,1*4(%%eax)\n" );
 879     output( "\tmovl 0xb0(%%ecx),%%edx\n" ); /* Eax */
 880     output( "\t.byte 0x36\n\tmovl %%edx,0*4(%%eax)\n" );
 881
 882     output( "\tpushl 0x98(%%ecx)\n" );      /* SegDs */
 883
 884     output( "\tmovl 0x9c(%%ecx),%%edi\n" ); /* Edi */
 885     output( "\tmovl 0xa0(%%ecx),%%esi\n" ); /* Esi */
 886     output( "\tmovl 0xa4(%%ecx),%%ebx\n" ); /* Ebx */
 887     output( "\tmovl 0xa8(%%ecx),%%edx\n" ); /* Edx */
 888     output( "\tmovl 0xb4(%%ecx),%%ebp\n" ); /* Ebp */
 889     output( "\tmovl 0xac(%%ecx),%%ecx\n" ); /* Ecx */
 890
 891     output( "\tpopl %%ds\n" );
 892     output( "\tmovl %%eax,%%esp\n" );
 893
 894     output( "\tpopl %%eax\n" );
 895     output( "\tiret\n" );
 896
 897     output("3:\n");
 898
 899     /* Restore the context when the stack segment changes. We can't use
 900      * the same code as above because we do not know if the stack segment
 901      * is 16 or 32 bit, and 'movl' will throw an exception when we try to
 902      * access memory above the limit. */
 903
 904     output( "\tmovl 0x9c(%%ecx),%%edi\n" ); /* Edi */
 905     output( "\tmovl 0xa0(%%ecx),%%esi\n" ); /* Esi */
 906     output( "\tmovl 0xa4(%%ecx),%%ebx\n" ); /* Ebx */
 907     output( "\tmovl 0xa8(%%ecx),%%edx\n" ); /* Edx */
 908     output( "\tmovl 0xb0(%%ecx),%%eax\n" ); /* Eax */
 909     output( "\tmovl 0xb4(%%ecx),%%ebp\n" ); /* Ebp */
 910
 911     output( "\tpushl 0xc8(%%ecx)\n" );      /* SegSs */
 912     output( "\tpopl %%ss\n" );
 913     output( "\tmovl 0xc4(%%ecx),%%esp\n" ); /* Esp */
 914
 915     output( "\tpushl 0xc0(%%ecx)\n" );      /* EFlags */
 916     output( "\tpushl 0xbc(%%ecx)\n" );      /* SegCs */
 917     output( "\tpushl 0xb8(%%ecx)\n" );      /* Eip */
 918     output( "\tpushl 0x98(%%ecx)\n" );      /* SegDs */
 919     output( "\tmovl 0xac(%%ecx),%%ecx\n" ); /* Ecx */
 920
 921     output( "\tpopl %%ds\n" );
 922     output( "\tiret\n" );
 923     output_cfi( ".cfi_endproc" );
 924     output_function_size( "__wine_call_from_regs" );
 925
 926     function_header( "__wine_restore_regs" );
 927     output_cfi( ".cfi_startproc" );
 928     output( "\tmovl 4(%%esp),%%ecx\n" );
 929     output( "\tjmp 2b\n" );
 930     output_cfi( ".cfi_endproc" );
 931     output_function_size( "__wine_restore_regs" );
 932 }
 933
 934
 935 /*******************************************************************
 936  *         BuildPendingEventCheck
 937  *
 938  * Build a function that checks whether there are any
 939  * pending DPMI events.
 940  *
 941  * Stack layout:
 942  *
 943  * (sp+12) long   eflags
 944  * (sp+6)  long   cs
 945  * (sp+2)  long   ip
 946  * (sp)    word   fs
 947  *
 948  * On entry to function, fs register points to a valid TEB.
 949  * On exit from function, stack will be popped.
 950  */
 951 static void BuildPendingEventCheck(void)
 952 {
 953     /* Function header */
 954
 955     function_header( "DPMI_PendingEventCheck" );
 956
 957     /* Check for pending events. */
 958
 959     output( "\t.byte 0x64\n\ttestl $0xffffffff,(%d)\n", VM86_PENDING_OFFSET );
 960     output( "\tje %s\n", asm_name("DPMI_PendingEventCheck_Cleanup") );
 961     output( "\t.byte 0x64\n\ttestl $0xffffffff,(%d)\n", DPMI_VIF_OFFSET );
 962     output( "\tje %s\n", asm_name("DPMI_PendingEventCheck_Cleanup") );
 963
 964     /* Process pending events. */
 965
 966     output( "\tsti\n" );
 967
 968     /* Start cleanup. Restore fs register. */
 969
 970     output( "%s\n", asm_globl("DPMI_PendingEventCheck_Cleanup") );
 971     output( "\tpopw %%fs\n" );
 972
 973     /* Return from function. */
 974
 975     output( "%s\n", asm_globl("DPMI_PendingEventCheck_Return") );
 976     output( "\tiret\n" );
 977
 978     output_function_size( "DPMI_PendingEventCheck" );
 979 }
 980
 981
 982 /*******************************************************************
 983  *         output_asm_relays16
 984  *
 985  * Build all the 16-bit relay callbacks
 986  */
 987 void output_asm_relays16(void)
 988 {
 989     /* File header */
 990
 991     output( "\t.text\n" );
 992     output( "%s:\n\n", asm_name("__wine_spec_thunk_text_16") );
 993
 994     output( "%s\n", asm_globl("__wine_call16_start") );
 995
 996     /* Standard CallFrom16 routine */
 997     BuildCallFrom16Core( 0, 0 );
 998
 999     /* Register CallFrom16 routine */
1000     BuildCallFrom16Core( 1, 0 );
1001
1002     /* C16ThkSL CallFrom16 routine */
1003     BuildCallFrom16Core( 0, 1 );
1004
1005     /* Standard CallTo16 routine */
1006     BuildCallTo16Core( 0 );
1007
1008     /* Register CallTo16 routine */
1009     BuildCallTo16Core( 1 );
1010
1011     /* Standard CallTo16 return stub */
1012     BuildRet16Func();
1013
1014     /* CBClientThunkSL routine */
1015     BuildCallTo32CBClient( 0 );
1016
1017     /* CBClientThunkSLEx routine */
1018     BuildCallTo32CBClient( 1  );
1019
1020     /* Pending DPMI events check stub */
1021     BuildPendingEventCheck();
1022
1023     output( "%s\n", asm_globl("__wine_call16_end") );
1024     output_function_size( "__wine_spec_thunk_text_16" );
1025
1026     /* Declare the return address and data selector variables */
1027     output( "\n\t.data\n\t.align %d\n", get_alignment(4) );
1028     output( "%s\n\t.long 0\n", asm_globl("CallTo16_DataSelector") );
1029     output( "%s\n\t.long 0\n", asm_globl("CallTo16_TebSelector") );
1030
1031     output( "\t.text\n" );
1032     output( "%s:\n", asm_name("__wine_spec_thunk_text_32") );
1033     build_call_from_regs_x86();
1034     output_function_size( "__wine_spec_thunk_text_32" );
1035 }
1036
1037
1038 /*******************************************************************
1039  *         output_asm_relays
1040  *
1041  * Build all the assembly relay callbacks
1042  */
1043 void output_asm_relays(void)
1044 {
1045     switch (target_cpu)
1046     {
1047     case CPU_x86:
1048         build_call_from_regs_x86();
1049         break;
1050     default:
1051         break;
1052     }
1053 }