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,SystemReserved2) + 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         return;
 228     }
 229
 230
 231     /* Build register CONTEXT */
 232     if ( reg_func )
 233     {
 234         output( "\tsubl $0x2cc,%%esp\n" );       /* sizeof(CONTEXT86) */
 235
 236         output( "\tmovl %%ecx,0xc0(%%esp)\n" );  /* EFlags */
 237
 238         output( "\tmovl %%eax,0xb0(%%esp)\n" );  /* Eax */
 239         output( "\tmovl %%ebx,0xa4(%%esp)\n" );  /* Ebx */
 240         output( "\tmovl %%esi,0xa0(%%esp)\n" );  /* Esi */
 241         output( "\tmovl %%edi,0x9c(%%esp)\n" );  /* Edi */
 242
 243         output( "\tmovl 0x0c(%%edx),%%eax\n");   /* FIELD_OFFSET(STACK16FRAME,ebp) */
 244         output( "\tmovl %%eax,0xb4(%%esp)\n" );  /* Ebp */
 245         output( "\tmovl 0x08(%%edx),%%eax\n");   /* FIELD_OFFSET(STACK16FRAME,ecx) */
 246         output( "\tmovl %%eax,0xac(%%esp)\n" );  /* Ecx */
 247         output( "\tmovl 0x04(%%edx),%%eax\n");   /* FIELD_OFFSET(STACK16FRAME,edx) */
 248         output( "\tmovl %%eax,0xa8(%%esp)\n" );  /* Edx */
 249
 250         output( "\tmovzwl 0x10(%%edx),%%eax\n"); /* FIELD_OFFSET(STACK16FRAME,ds) */
 251         output( "\tmovl %%eax,0x98(%%esp)\n" );  /* SegDs */
 252         output( "\tmovzwl 0x12(%%edx),%%eax\n"); /* FIELD_OFFSET(STACK16FRAME,es) */
 253         output( "\tmovl %%eax,0x94(%%esp)\n" );  /* SegEs */
 254         output( "\tmovzwl 0x14(%%edx),%%eax\n"); /* FIELD_OFFSET(STACK16FRAME,fs) */
 255         output( "\tmovl %%eax,0x90(%%esp)\n" );  /* SegFs */
 256         output( "\tmovzwl 0x16(%%edx),%%eax\n"); /* FIELD_OFFSET(STACK16FRAME,gs) */
 257         output( "\tmovl %%eax,0x8c(%%esp)\n" );  /* SegGs */
 258
 259         output( "\tmovzwl 0x2e(%%edx),%%eax\n"); /* FIELD_OFFSET(STACK16FRAME,cs) */
 260         output( "\tmovl %%eax,0xbc(%%esp)\n" );  /* SegCs */
 261         output( "\tmovzwl 0x2c(%%edx),%%eax\n"); /* FIELD_OFFSET(STACK16FRAME,ip) */
 262         output( "\tmovl %%eax,0xb8(%%esp)\n" );  /* Eip */
 263
 264         output( "\t.byte 0x64\n\tmovzwl (%d), %%eax\n", STACKOFFSET+2 );
 265         output( "\tmovl %%eax,0xc8(%%esp)\n" );  /* SegSs */
 266         output( "\t.byte 0x64\n\tmovzwl (%d), %%eax\n", STACKOFFSET );
 267         output( "\taddl $0x2c,%%eax\n");         /* FIELD_OFFSET(STACK16FRAME,ip) */
 268         output( "\tmovl %%eax,0xc4(%%esp)\n" );  /* Esp */
 269 #if 0
 270         output( "\tfsave 0x1c(%%esp)\n" ); /* FloatSave */
 271 #endif
 272
 273         /* Push address of CONTEXT86 structure -- popped by the relay routine */
 274         output( "\tmovl %%esp,%%eax\n" );
 275         output( "\tandl $~15,%%esp\n" );
 276         output( "\tsubl $4,%%esp\n" );
 277         output( "\tpushl %%eax\n" );
 278     }
 279     else
 280     {
 281         output( "\tsubl $8,%%esp\n" );
 282         output( "\tandl $~15,%%esp\n" );
 283         output( "\taddl $8,%%esp\n" );
 284     }
 285
 286     /* Call relay routine (which will call the API entry point) */
 287     output( "\tleal 0x30(%%edx),%%eax\n" ); /* sizeof(STACK16FRAME) */
 288     output( "\tpushl %%eax\n" );
 289     output( "\tpushl 0x26(%%edx)\n");  /* FIELD_OFFSET(STACK16FRAME,entry_point) */
 290     output( "\tcall *0x20(%%edx)\n");  /* FIELD_OFFSET(STACK16FRAME,relay) */
 291
 292     if ( reg_func )
 293     {
 294         output( "\tleal -748(%%ebp),%%ebx\n" ); /* sizeof(CONTEXT) + FIELD_OFFSET(STACK32FRAME,ebp) */
 295
 296         /* Switch stack back */
 297         output( "\t.byte 0x64\n\tmovw (%d), %%ss\n", STACKOFFSET+2 );
 298         output( "\t.byte 0x64\n\tmovzwl (%d), %%esp\n", STACKOFFSET );
 299         output( "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 300
 301         /* Get return address to CallFrom16 stub */
 302         output( "\taddw $0x14,%%sp\n" ); /* FIELD_OFFSET(STACK16FRAME,callfrom_ip)-4 */
 303         output( "\tpopl %%eax\n" );
 304         output( "\tpopl %%edx\n" );
 305
 306         /* Restore all registers from CONTEXT */
 307         output( "\tmovw 0xc8(%%ebx),%%ss\n");   /* SegSs */
 308         output( "\tmovl 0xc4(%%ebx),%%esp\n");  /* Esp */
 309         output( "\taddl $4, %%esp\n" );  /* room for final return address */
 310
 311         output( "\tpushw 0xbc(%%ebx)\n");  /* SegCs */
 312         output( "\tpushw 0xb8(%%ebx)\n");  /* Eip */
 313         output( "\tpushl %%edx\n" );
 314         output( "\tpushl %%eax\n" );
 315         output( "\tpushl 0xc0(%%ebx)\n");  /* EFlags */
 316         output( "\tpushl 0x98(%%ebx)\n");  /* SegDs */
 317
 318         output( "\tpushl 0x94(%%ebx)\n");  /* SegEs */
 319         output( "\tpopl %%es\n" );
 320         output( "\tpushl 0x90(%%ebx)\n");  /* SegFs */
 321         output( "\tpopl %%fs\n" );
 322         output( "\tpushl 0x8c(%%ebx)\n");  /* SegGs */
 323         output( "\tpopl %%gs\n" );
 324
 325         output( "\tmovl 0xb4(%%ebx),%%ebp\n");  /* Ebp */
 326         output( "\tmovl 0xa0(%%ebx),%%esi\n");  /* Esi */
 327         output( "\tmovl 0x9c(%%ebx),%%edi\n");  /* Edi */
 328         output( "\tmovl 0xb0(%%ebx),%%eax\n");  /* Eax */
 329         output( "\tmovl 0xa8(%%ebx),%%edx\n");  /* Edx */
 330         output( "\tmovl 0xac(%%ebx),%%ecx\n");  /* Ecx */
 331         output( "\tmovl 0xa4(%%ebx),%%ebx\n");  /* Ebx */
 332
 333         output( "\tpopl %%ds\n" );
 334         output( "\tpopfl\n" );
 335         output( "\tlret\n" );
 336     }
 337     else
 338     {
 339         /* Switch stack back */
 340         output( "\t.byte 0x64\n\tmovw (%d), %%ss\n", STACKOFFSET+2 );
 341         output( "\t.byte 0x64\n\tmovzwl (%d), %%esp\n", STACKOFFSET );
 342         output( "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 343
 344         /* Restore registers */
 345         output( "\tpopl %%edx\n" );
 346         output( "\tpopl %%ecx\n" );
 347         output( "\tpopl %%ebp\n" );
 348         output( "\tpopw %%ds\n" );
 349         output( "\tpopw %%es\n" );
 350         output( "\tpopw %%fs\n" );
 351         output( "\tpopw %%gs\n" );
 352
 353         /* Return to return stub which will return to caller */
 354         output( "\tlret $12\n" );
 355     }
 356     if (thunk) output_function_size( "__wine_call_from_16_thunk" );
 357     else if (reg_func) output_function_size( "__wine_call_from_16_regs" );
 358     else output_function_size( "__wine_call_from_16" );
 359 }
 360
 361
 362 /*******************************************************************
 363  *         BuildCallTo16Core
 364  *
 365  * This routine builds the core routines used in 32->16 thunks:
 366  *
 367  * extern DWORD WINAPI wine_call_to_16( FARPROC16 target, DWORD cbArgs, PEXCEPTION_HANDLER handler );
 368  * extern void WINAPI wine_call_to_16_regs( CONTEXT86 *context, DWORD cbArgs, PEXCEPTION_HANDLER handler );
 369  *
 370  * These routines can be called directly from 32-bit code.
 371  *
 372  * All routines expect that the 16-bit stack contents (arguments) and the
 373  * return address (segptr to CallTo16_Ret) were already set up by the
 374  * caller; nb_args must contain the number of bytes to be conserved.  The
 375  * 16-bit SS:SP will be set accordingly.
 376  *
 377  * All other registers are either taken from the CONTEXT86 structure
 378  * or else set to default values.  The target routine address is either
 379  * given directly or taken from the CONTEXT86.
 380  */
 381 static void BuildCallTo16Core( int reg_func )
 382 {
 383     const char *name = reg_func ? "wine_call_to_16_regs" : "wine_call_to_16";
 384
 385     /* Function header */
 386     function_header( name );
 387
 388     /* Function entry sequence */
 389     output( "\tpushl %%ebp\n" );
 390     output( "\tmovl %%esp, %%ebp\n" );
 391
 392     /* Save the 32-bit registers */
 393     output( "\tpushl %%ebx\n" );
 394     output( "\tpushl %%esi\n" );
 395     output( "\tpushl %%edi\n" );
 396     output( "\t.byte 0x64\n\tmov %%gs,(%d)\n", GS_OFFSET );
 397
 398     /* Setup exception frame */
 399     output( "\t.byte 0x64\n\tpushl (%d)\n", STACKOFFSET );
 400     output( "\tpushl 16(%%ebp)\n" ); /* handler */
 401     output( "\t.byte 0x64\n\tpushl (0)\n" );
 402     output( "\t.byte 0x64\n\tmovl %%esp,(0)\n" );
 403
 404     /* Call the actual CallTo16 routine (simulate a lcall) */
 405     output( "\tpushl %%cs\n" );
 406     output( "\tcall .L%s\n", name );
 407
 408     /* Remove exception frame */
 409     output( "\t.byte 0x64\n\tpopl (0)\n" );
 410     output( "\taddl $4, %%esp\n" );
 411     output( "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 412
 413     if ( !reg_func )
 414     {
 415         /* Convert return value */
 416         output( "\tandl $0xffff,%%eax\n" );
 417         output( "\tshll $16,%%edx\n" );
 418         output( "\torl %%edx,%%eax\n" );
 419     }
 420     else
 421     {
 422         /*
 423          * Modify CONTEXT86 structure to contain new values
 424          *
 425          * NOTE:  We restore only EAX, EBX, EDX, EDX, EBP, and ESP.
 426          *        The segment registers as well as ESI and EDI should
 427          *        not be modified by a well-behaved 16-bit routine in
 428          *        any case.  [If necessary, we could restore them as well,
 429          *        at the cost of a somewhat less efficient return path.]
 430          */
 431
 432         output( "\tmovl 0x14(%%esp),%%edi\n" ); /* FIELD_OFFSET(STACK32FRAME,target) - FIELD_OFFSET(STACK32FRAME,edi) */
 433                 /* everything above edi has been popped already */
 434
 435         output( "\tmovl %%eax,0xb0(%%edi)\n");  /* Eax */
 436         output( "\tmovl %%ebx,0xa4(%%edi)\n");  /* Ebx */
 437         output( "\tmovl %%ecx,0xac(%%edi)\n");  /* Ecx */
 438         output( "\tmovl %%edx,0xa8(%%edi)\n");  /* Edx */
 439         output( "\tmovl %%ebp,0xb4(%%edi)\n");  /* Ebp */
 440         output( "\tmovl %%esi,0xc4(%%edi)\n");  /* Esp */
 441                  /* The return glue code saved %esp into %esi */
 442     }
 443
 444     /* Restore the 32-bit registers */
 445     output( "\tpopl %%edi\n" );
 446     output( "\tpopl %%esi\n" );
 447     output( "\tpopl %%ebx\n" );
 448
 449     /* Function exit sequence */
 450     output( "\tpopl %%ebp\n" );
 451     output( "\tret $12\n" );
 452
 453
 454     /* Start of the actual CallTo16 routine */
 455
 456     output( ".L%s:\n", name );
 457
 458     /* Switch to the 16-bit stack */
 459     output( "\tmovl %%esp,%%edx\n" );
 460     output( "\t.byte 0x64\n\tmovw (%d),%%ss\n", STACKOFFSET + 2);
 461     output( "\t.byte 0x64\n\tmovw (%d),%%sp\n", STACKOFFSET );
 462     output( "\t.byte 0x64\n\tmovl %%edx,(%d)\n", STACKOFFSET );
 463
 464     /* Make %bp point to the previous stackframe (built by CallFrom16) */
 465     output( "\tmovzwl %%sp,%%ebp\n" );
 466     output( "\tleal 0x2a(%%ebp),%%ebp\n");  /* FIELD_OFFSET(STACK16FRAME,bp) */
 467
 468     /* Add the specified offset to the new sp */
 469     output( "\tsubw 0x2c(%%edx), %%sp\n");  /* FIELD_OFFSET(STACK32FRAME,nb_args) */
 470
 471     if (reg_func)
 472     {
 473         /* Push the called routine address */
 474         output( "\tmovl 0x28(%%edx),%%edx\n");  /* FIELD_OFFSET(STACK32FRAME,target) */
 475         output( "\tpushw 0xbc(%%edx)\n");  /* SegCs */
 476         output( "\tpushw 0xb8(%%edx)\n");  /* Eip */
 477
 478         /* Get the registers */
 479         output( "\tpushw 0x98(%%edx)\n");  /* SegDs */
 480         output( "\tpushl 0x94(%%edx)\n");  /* SegEs */
 481         output( "\tpopl %%es\n" );
 482         output( "\tpushl 0x90(%%edx)\n");  /* SegFs */
 483         output( "\tpopl %%fs\n" );
 484         output( "\tpushl 0x8c(%%edx)\n");  /* SegGs */
 485         output( "\tpopl %%gs\n" );
 486         output( "\tmovl 0xb4(%%edx),%%ebp\n");  /* Ebp */
 487         output( "\tmovl 0xa0(%%edx),%%esi\n");  /* Esi */
 488         output( "\tmovl 0x9c(%%edx),%%edi\n");  /* Edi */
 489         output( "\tmovl 0xb0(%%edx),%%eax\n");  /* Eax */
 490         output( "\tmovl 0xa4(%%edx),%%ebx\n");  /* Ebx */
 491         output( "\tmovl 0xac(%%edx),%%ecx\n");  /* Ecx */
 492         output( "\tmovl 0xa8(%%edx),%%edx\n");  /* Edx */
 493
 494         /* Get the 16-bit ds */
 495         output( "\tpopw %%ds\n" );
 496     }
 497     else  /* not a register function */
 498     {
 499         /* Push the called routine address */
 500         output( "\tpushl 0x28(%%edx)\n"); /* FIELD_OFFSET(STACK32FRAME,target) */
 501
 502         /* Set %fs and %gs to the value saved by the last CallFrom16 */
 503         output( "\tpushw -22(%%ebp)\n" ); /* FIELD_OFFSET(STACK16FRAME,fs)-FIELD_OFFSET(STACK16FRAME,bp) */
 504         output( "\tpopw %%fs\n" );
 505         output( "\tpushw -20(%%ebp)\n" ); /* FIELD_OFFSET(STACK16FRAME,gs)-FIELD_OFFSET(STACK16FRAME,bp) */
 506         output( "\tpopw %%gs\n" );
 507
 508         /* Set %ds and %es (and %ax just in case) equal to %ss */
 509         output( "\tmovw %%ss,%%ax\n" );
 510         output( "\tmovw %%ax,%%ds\n" );
 511         output( "\tmovw %%ax,%%es\n" );
 512     }
 513
 514     /* Jump to the called routine */
 515     output( "\t.byte 0x66\n" );
 516     output( "\tlret\n" );
 517
 518     /* Function footer */
 519     output_function_size( name );
 520 }
 521
 522
 523 /*******************************************************************
 524  *         BuildRet16Func
 525  *
 526  * Build the return code for 16-bit callbacks
 527  */
 528 static void BuildRet16Func(void)
 529 {
 530     function_header( "__wine_call_to_16_ret" );
 531
 532     /* Save %esp into %esi */
 533     output( "\tmovl %%esp,%%esi\n" );
 534
 535     /* Restore 32-bit segment registers */
 536
 537     output( "\t.byte 0x2e\n\tmovl %s", asm_name("CallTo16_DataSelector") );
 538     output( "-%s,%%edi\n", asm_name("__wine_call16_start") );
 539     output( "\tmovw %%di,%%ds\n" );
 540     output( "\tmovw %%di,%%es\n" );
 541
 542     output( "\t.byte 0x2e\n\tmov %s", asm_name("CallTo16_TebSelector") );
 543     output( "-%s,%%fs\n", asm_name("__wine_call16_start") );
 544
 545     output( "\t.byte 0x64\n\tmov (%d),%%gs\n", GS_OFFSET );
 546
 547     /* Restore the 32-bit stack */
 548
 549     output( "\tmovw %%di,%%ss\n" );
 550     output( "\t.byte 0x64\n\tmovl (%d),%%esp\n", STACKOFFSET );
 551
 552     /* Return to caller */
 553
 554     output( "\tlret\n" );
 555     output_function_size( "__wine_call_to_16_ret" );
 556 }
 557
 558
 559 /*******************************************************************
 560  *         BuildCallTo32CBClient
 561  *
 562  * Call a CBClient relay stub from 32-bit code (KERNEL.620).
 563  *
 564  * Since the relay stub is itself 32-bit, this should not be a problem;
 565  * unfortunately, the relay stubs are expected to switch back to a
 566  * 16-bit stack (and 16-bit code) after completion :-(
 567  *
 568  * This would conflict with our 16- vs. 32-bit stack handling, so
 569  * we simply switch *back* to our 32-bit stack before returning to
 570  * the caller ...
 571  *
 572  * The CBClient relay stub expects to be called with the following
 573  * 16-bit stack layout, and with ebp and ebx pointing into the 16-bit
 574  * stack at the designated places:
 575  *
 576  *    ...
 577  *  (ebp+14) original arguments to the callback routine
 578  *  (ebp+10) far return address to original caller
 579  *  (ebp+6)  Thunklet target address
 580  *  (ebp+2)  Thunklet relay ID code
 581  *  (ebp)    BP (saved by CBClientGlueSL)
 582  *  (ebp-2)  SI (saved by CBClientGlueSL)
 583  *  (ebp-4)  DI (saved by CBClientGlueSL)
 584  *  (ebp-6)  DS (saved by CBClientGlueSL)
 585  *
 586  *   ...     buffer space used by the 16-bit side glue for temp copies
 587  *
 588  *  (ebx+4)  far return address to 16-bit side glue code
 589  *  (ebx)    saved 16-bit ss:sp (pointing to ebx+4)
 590  *
 591  * The 32-bit side glue code accesses both the original arguments (via ebp)
 592  * and the temporary copies prepared by the 16-bit side glue (via ebx).
 593  * After completion, the stub will load ss:sp from the buffer at ebx
 594  * and perform a far return to 16-bit code.
 595  *
 596  * To trick the relay stub into returning to us, we replace the 16-bit
 597  * return address to the glue code by a cs:ip pair pointing to our
 598  * return entry point (the original return address is saved first).
 599  * Our return stub thus called will then reload the 32-bit ss:esp and
 600  * return to 32-bit code (by using and ss:esp value that we have also
 601  * pushed onto the 16-bit stack before and a cs:eip values found at
 602  * that position on the 32-bit stack).  The ss:esp to be restored is
 603  * found relative to the 16-bit stack pointer at:
 604  *
 605  *  (ebx-4)   ss  (flat)
 606  *  (ebx-8)   sp  (32-bit stack pointer)
 607  *
 608  * The second variant of this routine, CALL32_CBClientEx, which is used
 609  * to implement KERNEL.621, has to cope with yet another problem: Here,
 610  * the 32-bit side directly returns to the caller of the CBClient thunklet,
 611  * restoring registers saved by CBClientGlueSL and cleaning up the stack.
 612  * As we have to return to our 32-bit code first, we have to adapt the
 613  * layout of our temporary area so as to include values for the registers
 614  * that are to be restored, and later (in the implementation of KERNEL.621)
 615  * we *really* restore them. The return stub restores DS, DI, SI, and BP
 616  * from the stack, skips the next 8 bytes (CBClient relay code / target),
 617  * and then performs a lret NN, where NN is the number of arguments to be
 618  * removed. Thus, we prepare our temporary area as follows:
 619  *
 620  *     (ebx+22) 16-bit cs  (this segment)
 621  *     (ebx+20) 16-bit ip  ('16-bit' return entry point)
 622  *     (ebx+16) 32-bit ss  (flat)
 623  *     (ebx+12) 32-bit sp  (32-bit stack pointer)
 624  *     (ebx+10) 16-bit bp  (points to ebx+24)
 625  *     (ebx+8)  16-bit si  (ignored)
 626  *     (ebx+6)  16-bit di  (ignored)
 627  *     (ebx+4)  16-bit ds  (we actually use the flat DS here)
 628  *     (ebx+2)  16-bit ss  (16-bit stack segment)
 629  *     (ebx+0)  16-bit sp  (points to ebx+4)
 630  *
 631  * Note that we ensure that DS is not changed and remains the flat segment,
 632  * and the 32-bit stack pointer our own return stub needs fits just
 633  * perfectly into the 8 bytes that are skipped by the Windows stub.
 634  * One problem is that we have to determine the number of removed arguments,
 635  * as these have to be really removed in KERNEL.621. Thus, the BP value
 636  * that we place in the temporary area to be restored, contains the value
 637  * that SP would have if no arguments were removed. By comparing the actual
 638  * value of SP with this value in our return stub we can compute the number
 639  * of removed arguments. This is then returned to KERNEL.621.
 640  *
 641  * The stack layout of this function:
 642  * (ebp+20)  nArgs     pointer to variable receiving nr. of args (Ex only)
 643  * (ebp+16)  esi       pointer to caller's esi value
 644  * (ebp+12)  arg       ebp value to be set for relay stub
 645  * (ebp+8)   func      CBClient relay stub address
 646  * (ebp+4)   ret addr
 647  * (ebp)     ebp
 648  */
 649 static void BuildCallTo32CBClient( int isEx )
 650 {
 651     function_header( isEx ? "CALL32_CBClientEx" : "CALL32_CBClient" );
 652
 653     /* Entry code */
 654
 655     output( "\tpushl %%ebp\n" );
 656     output( "\tmovl %%esp,%%ebp\n" );
 657     output( "\tpushl %%edi\n" );
 658     output( "\tpushl %%esi\n" );
 659     output( "\tpushl %%ebx\n" );
 660
 661     /* Get pointer to temporary area and save the 32-bit stack pointer */
 662
 663     output( "\tmovl 16(%%ebp), %%ebx\n" );
 664     output( "\tleal -8(%%esp), %%eax\n" );
 665
 666     if ( !isEx )
 667         output( "\tmovl %%eax, -8(%%ebx)\n" );
 668     else
 669         output( "\tmovl %%eax, 12(%%ebx)\n" );
 670
 671     /* Set up registers and call CBClient relay stub (simulating a far call) */
 672
 673     output( "\tmovl 20(%%ebp), %%esi\n" );
 674     output( "\tmovl (%%esi), %%esi\n" );
 675
 676     output( "\tmovl 8(%%ebp), %%eax\n" );
 677     output( "\tmovl 12(%%ebp), %%ebp\n" );
 678
 679     output( "\tpushl %%cs\n" );
 680     output( "\tcall *%%eax\n" );
 681
 682     /* Return new esi value to caller */
 683
 684     output( "\tmovl 32(%%esp), %%edi\n" );
 685     output( "\tmovl %%esi, (%%edi)\n" );
 686
 687     /* Return argument size to caller */
 688     if ( isEx )
 689     {
 690         output( "\tmovl 36(%%esp), %%ebx\n" );
 691         output( "\tmovl %%ebp, (%%ebx)\n" );
 692     }
 693
 694     /* Restore registers and return */
 695
 696     output( "\tpopl %%ebx\n" );
 697     output( "\tpopl %%esi\n" );
 698     output( "\tpopl %%edi\n" );
 699     output( "\tpopl %%ebp\n" );
 700     output( "\tret\n" );
 701     output_function_size( isEx ? "CALL32_CBClientEx" : "CALL32_CBClient" );
 702
 703     /* '16-bit' return stub */
 704
 705     function_header( isEx ? "CALL32_CBClientEx_Ret" : "CALL32_CBClient_Ret" );
 706     if ( !isEx )
 707     {
 708         output( "\tmovzwl %%sp, %%ebx\n" );
 709         output( "\tlssl %%ss:-16(%%ebx), %%esp\n" );
 710     }
 711     else
 712     {
 713         output( "\tmovzwl %%bp, %%ebx\n" );
 714         output( "\tsubw %%bp, %%sp\n" );
 715         output( "\tmovzwl %%sp, %%ebp\n" );
 716         output( "\tlssl %%ss:-12(%%ebx), %%esp\n" );
 717     }
 718     output( "\tlret\n" );
 719     output_function_size( isEx ? "CALL32_CBClientEx_Ret" : "CALL32_CBClient_Ret" );
 720 }
 721
 722
 723 /*******************************************************************
 724  *         BuildCallFrom32Regs
 725  *
 726  * Build a 32-bit-to-Wine call-back function for a 'register' function.
 727  * 'args' is the number of dword arguments.
 728  *
 729  * Stack layout:
 730  *   ...
 731  * (ebp+20)  first arg
 732  * (ebp+16)  ret addr to user code
 733  * (ebp+12)  func to call (relative to relay code ret addr)
 734  * (ebp+8)   number of args
 735  * (ebp+4)   ret addr to relay code
 736  * (ebp+0)   saved ebp
 737  * (ebp-128) buffer area to allow stack frame manipulation
 738  * (ebp-332) CONTEXT86 struct
 739  * (ebp-336) padding for stack alignment
 740  * (ebp-336-n) CONTEXT86 *argument
 741  *  ....     other arguments copied from (ebp+12)
 742  *
 743  * The entry point routine is called with a CONTEXT* extra argument,
 744  * following the normal args. In this context structure, EIP_reg
 745  * contains the return address to user code, and ESP_reg the stack
 746  * pointer on return (with the return address and arguments already
 747  * removed).
 748  */
 749 static void BuildCallFrom32Regs(void)
 750 {
 751     static const int STACK_SPACE = 128 + 0x2cc /* sizeof(CONTEXT86) */;
 752
 753     /* Function header */
 754
 755     function_header( "__wine_call_from_32_regs" );
 756
 757     /* Allocate some buffer space on the stack */
 758
 759     output( "\tpushl %%ebp\n" );
 760     output( "\tmovl %%esp,%%ebp\n ");
 761     output( "\tleal -%d(%%esp),%%esp\n", STACK_SPACE );
 762
 763     /* Build the context structure */
 764
 765     output( "\tmovl %%eax,0xb0(%%esp)\n" );  /* Eax */
 766     output( "\tpushfl\n" );
 767     output( "\tpopl %%eax\n" );
 768     output( "\tmovl %%eax,0xc0(%%esp)\n");  /* EFlags */
 769     output( "\tmovl 0(%%ebp),%%eax\n" );
 770     output( "\tmovl %%eax,0xb4(%%esp)\n");  /* Ebp */
 771     output( "\tmovl %%ebx,0xa4(%%esp)\n");  /* Ebx */
 772     output( "\tmovl %%ecx,0xac(%%esp)\n");  /* Ecx */
 773     output( "\tmovl %%edx,0xa8(%%esp)\n");  /* Edx */
 774     output( "\tmovl %%esi,0xa0(%%esp)\n");  /* Esi */
 775     output( "\tmovl %%edi,0x9c(%%esp)\n");  /* Edi */
 776
 777     output( "\txorl %%eax,%%eax\n" );
 778     output( "\tmovw %%cs,%%ax\n" );
 779     output( "\tmovl %%eax,0xbc(%%esp)\n");  /* SegCs */
 780     output( "\tmovw %%es,%%ax\n" );
 781     output( "\tmovl %%eax,0x94(%%esp)\n");  /* SegEs */
 782     output( "\tmovw %%fs,%%ax\n" );
 783     output( "\tmovl %%eax,0x90(%%esp)\n");  /* SegFs */
 784     output( "\tmovw %%gs,%%ax\n" );
 785     output( "\tmovl %%eax,0x8c(%%esp)\n");  /* SegGs */
 786     output( "\tmovw %%ss,%%ax\n" );
 787     output( "\tmovl %%eax,0xc8(%%esp)\n");  /* SegSs */
 788     output( "\tmovw %%ds,%%ax\n" );
 789     output( "\tmovl %%eax,0x98(%%esp)\n");  /* SegDs */
 790     output( "\tmovw %%ax,%%es\n" );  /* set %es equal to %ds just in case */
 791
 792     output( "\tmovl $0x10007,0(%%esp)\n");  /* ContextFlags */
 793
 794     output( "\tmovl 16(%%ebp),%%eax\n" ); /* Get %eip at time of call */
 795     output( "\tmovl %%eax,0xb8(%%esp)\n");  /* Eip */
 796
 797     /* Transfer the arguments */
 798
 799     output( "\tmovl 8(%%ebp),%%ecx\n" );    /* fetch number of args to copy */
 800     output( "\tleal 4(,%%ecx,4),%%edx\n" ); /* add 4 for context arg */
 801     output( "\tsubl %%edx,%%esp\n" );
 802     output( "\tandl $~15,%%esp\n" );
 803     output( "\tleal 20(%%ebp),%%esi\n" );  /* get %esp at time of call */
 804     output( "\tmovl %%esp,%%edi\n" );
 805     output( "\ttest %%ecx,%%ecx\n" );
 806     output( "\tjz 1f\n" );
 807     output( "\tcld\n" );
 808     output( "\trep\n\tmovsl\n" );  /* copy args */
 809     output( "1:\tleal %d(%%ebp),%%eax\n", -STACK_SPACE );  /* get addr of context struct */
 810     output( "\tmovl %%eax,(%%edi)\n" );    /* and pass it as extra arg */
 811     output( "\tmovl %%esi,%d(%%ebp)\n", 0xc4 /* Esp */ - STACK_SPACE );
 812
 813     /* Call the entry point */
 814
 815     output( "\tmovl 4(%%ebp),%%eax\n" );   /* get relay code addr */
 816     output( "\taddl 12(%%ebp),%%eax\n" );
 817     output( "\tcall *%%eax\n" );
 818     output( "\tleal -%d(%%ebp),%%ecx\n", STACK_SPACE );
 819
 820     /* Restore the context structure */
 821
 822     output( "2:\tpushl 0x94(%%ecx)\n");     /* SegEs */
 823     output( "\tpopl %%es\n" );
 824     output( "\tpushl 0x90(%%ecx)\n");       /* SegFs */
 825     output( "\tpopl %%fs\n" );
 826     output( "\tpushl 0x8c(%%ecx)\n");       /* SegGs */
 827     output( "\tpopl %%gs\n" );
 828
 829     output( "\tmovl 0x9c(%%ecx),%%edi\n");  /* Edi */
 830     output( "\tmovl 0xa0(%%ecx),%%esi\n");  /* Esi */
 831     output( "\tmovl 0xa8(%%ecx),%%edx\n");  /* Edx */
 832     output( "\tmovl 0xa4(%%ecx),%%ebx\n");  /* Ebx */
 833     output( "\tmovl 0xb0(%%ecx),%%eax\n");  /* Eax */
 834     output( "\tmovl 0xb4(%%ecx),%%ebp\n");  /* Ebp */
 835
 836     output( "\tpushl 0xc8(%%ecx)\n");       /* SegSs */
 837     output( "\tpopl %%ss\n" );
 838     output( "\tmovl 0xc4(%%ecx),%%esp\n");  /* Esp */
 839
 840     output( "\tpushl 0xc0(%%ecx)\n");       /* EFlags */
 841     output( "\tpushl 0xbc(%%ecx)\n");       /* SegCs */
 842     output( "\tpushl 0xb8(%%ecx)\n");       /* Eip */
 843     output( "\tpushl 0x98(%%ecx)\n");       /* SegDs */
 844     output( "\tmovl 0xac(%%ecx),%%ecx\n");  /* Ecx */
 845
 846     output( "\tpopl %%ds\n" );
 847     output( "\tiret\n" );
 848     output_function_size( "__wine_call_from_32_regs" );
 849
 850     function_header( "__wine_call_from_32_restore_regs" );
 851     output( "\tmovl 4(%%esp),%%ecx\n" );
 852     output( "\tjmp 2b\n" );
 853     output_function_size( "__wine_call_from_32_restore_regs" );
 854 }
 855
 856
 857 /*******************************************************************
 858  *         BuildPendingEventCheck
 859  *
 860  * Build a function that checks whether there are any
 861  * pending DPMI events.
 862  *
 863  * Stack layout:
 864  *
 865  * (sp+12) long   eflags
 866  * (sp+6)  long   cs
 867  * (sp+2)  long   ip
 868  * (sp)    word   fs
 869  *
 870  * On entry to function, fs register points to a valid TEB.
 871  * On exit from function, stack will be popped.
 872  */
 873 static void BuildPendingEventCheck(void)
 874 {
 875     /* Function header */
 876
 877     function_header( "DPMI_PendingEventCheck" );
 878
 879     /* Check for pending events. */
 880
 881     output( "\t.byte 0x64\n\ttestl $0xffffffff,(%d)\n", VM86_PENDING_OFFSET );
 882     output( "\tje %s\n", asm_name("DPMI_PendingEventCheck_Cleanup") );
 883     output( "\t.byte 0x64\n\ttestl $0xffffffff,(%d)\n", DPMI_VIF_OFFSET );
 884     output( "\tje %s\n", asm_name("DPMI_PendingEventCheck_Cleanup") );
 885
 886     /* Process pending events. */
 887
 888     output( "\tsti\n" );
 889
 890     /* Start cleanup. Restore fs register. */
 891
 892     output( "%s\n", asm_globl("DPMI_PendingEventCheck_Cleanup") );
 893     output( "\tpopw %%fs\n" );
 894
 895     /* Return from function. */
 896
 897     output( "%s\n", asm_globl("DPMI_PendingEventCheck_Return") );
 898     output( "\tiret\n" );
 899
 900     output_function_size( "DPMI_PendingEventCheck" );
 901 }
 902
 903
 904 /*******************************************************************
 905  *         BuildRelays16
 906  *
 907  * Build all the 16-bit relay callbacks
 908  */
 909 void BuildRelays16(void)
 910 {
 911     if (target_cpu != CPU_x86)
 912     {
 913         output( "/* File not used with this architecture. Do not edit! */\n\n" );
 914         return;
 915     }
 916
 917     /* File header */
 918
 919     output( "/* File generated automatically. Do not edit! */\n\n" );
 920     output( "\t.text\n" );
 921
 922     output( "%s:\n\n", asm_name("__wine_spec_thunk_text_16") );
 923
 924     output( "%s\n", asm_globl("__wine_call16_start") );
 925
 926     /* Standard CallFrom16 routine */
 927     BuildCallFrom16Core( 0, 0 );
 928
 929     /* Register CallFrom16 routine */
 930     BuildCallFrom16Core( 1, 0 );
 931
 932     /* C16ThkSL CallFrom16 routine */
 933     BuildCallFrom16Core( 0, 1 );
 934
 935     /* Standard CallTo16 routine */
 936     BuildCallTo16Core( 0 );
 937
 938     /* Register CallTo16 routine */
 939     BuildCallTo16Core( 1 );
 940
 941     /* Standard CallTo16 return stub */
 942     BuildRet16Func();
 943
 944     /* CBClientThunkSL routine */
 945     BuildCallTo32CBClient( 0 );
 946
 947     /* CBClientThunkSLEx routine */
 948     BuildCallTo32CBClient( 1  );
 949
 950     /* Pending DPMI events check stub */
 951     BuildPendingEventCheck();
 952
 953     output( "%s\n", asm_globl("__wine_call16_end") );
 954     output_function_size( "__wine_spec_thunk_text_16" );
 955
 956     /* Declare the return address and data selector variables */
 957     output( "\n\t.data\n\t.align %d\n", get_alignment(4) );
 958     output( "%s\n\t.long 0\n", asm_globl("CallTo16_DataSelector") );
 959     output( "%s\n\t.long 0\n", asm_globl("CallTo16_TebSelector") );
 960     if (UsePIC) output( "wine_ldt_copy_ptr:\t.long %s\n", asm_name("wine_ldt_copy") );
 961
 962     output( "\t.text\n" );
 963     output( "%s:\n\n", asm_name("__wine_spec_thunk_text_32") );
 964     BuildCallFrom32Regs();
 965     output_function_size( "__wine_spec_thunk_text_32" );
 966
 967     output_gnu_stack_note();
 968 }
 969
 970
 971 /*******************************************************************
 972  *         build_call_from_regs_x86_64
 973  *
 974  * Build the register saving code for a 'register' entry point.
 975  *
 976  * Stack layout:
 977  *   ...
 978  * (rsp+16)  first arg
 979  * (rsp+8)   ret addr to user code
 980  * (rsp)     ret addr to relay code
 981  * (rsp-128) buffer area to allow stack frame manipulation
 982  *
 983  * Parameters:
 984  *  %rcx     number of args
 985  *  %rdx     entry point
 986  */
 987 static void build_call_from_regs_x86_64(void)
 988 {
 989     static const int STACK_SPACE = 128 + 0x4d0; /* size of x86_64 context */
 990
 991     /* Function header */
 992
 993     function_header( "__wine_call_from_regs" );
 994
 995     output( "\tsubq $%u,%%rsp\n", STACK_SPACE );
 996
 997     /* save registers into the context */
 998
 999     output( "\tmovq %%rax,0x78(%%rsp)\n" );
1000     output( "\tmovq %u(%%rsp),%%rax\n", STACK_SPACE + 16 );  /* saved %rcx on stack */
1001     output( "\tmovq %%rax,0x80(%%rsp)\n" );
1002     output( "\tmovq %u(%%rsp),%%rax\n", STACK_SPACE + 24 );  /* saved %rdx on stack */
1003     output( "\tmovq %%rax,0x88(%%rsp)\n" );
1004     output( "\tmovq %%rbx,0x90(%%rsp)\n" );
1005     output( "\tleaq %u(%%rsp),%%rax\n", STACK_SPACE + 16 );
1006     output( "\tmovq %%rax,0x98(%%rsp)\n" );
1007     output( "\tmovq %%rbp,0xa0(%%rsp)\n" );
1008     output( "\tmovq %%rsi,0xa8(%%rsp)\n" );
1009     output( "\tmovq %%rdi,0xb0(%%rsp)\n" );
1010     output( "\tmovq %%r8,0xb8(%%rsp)\n" );
1011     output( "\tmovq %%r9,0xc0(%%rsp)\n" );
1012     output( "\tmovq %%r10,0xc8(%%rsp)\n" );
1013     output( "\tmovq %%r11,0xd0(%%rsp)\n" );
1014     output( "\tmovq %%r12,0xd8(%%rsp)\n" );
1015     output( "\tmovq %%r13,0xe0(%%rsp)\n" );
1016     output( "\tmovq %%r14,0xe8(%%rsp)\n" );
1017     output( "\tmovq %%r15,0xf0(%%rsp)\n" );
1018     output( "\tmovq %u(%%rsp),%%rax\n", STACK_SPACE + 8 );
1019     output( "\tmovq %%rax,0xf8(%%rsp)\n" );
1020
1021     output( "\tstmxcsr 0x34(%%rsp)\n" );
1022     output( "\tfxsave 0x100(%%rsp)\n" );
1023     output( "\tmovdqa %%xmm0,0x1a0(%%rsp)\n" );
1024     output( "\tmovdqa %%xmm1,0x1b0(%%rsp)\n" );
1025     output( "\tmovdqa %%xmm2,0x1c0(%%rsp)\n" );
1026     output( "\tmovdqa %%xmm3,0x1d0(%%rsp)\n" );
1027     output( "\tmovdqa %%xmm4,0x1e0(%%rsp)\n" );
1028     output( "\tmovdqa %%xmm5,0x1f0(%%rsp)\n" );
1029     output( "\tmovdqa %%xmm6,0x200(%%rsp)\n" );
1030     output( "\tmovdqa %%xmm7,0x210(%%rsp)\n" );
1031     output( "\tmovdqa %%xmm8,0x220(%%rsp)\n" );
1032     output( "\tmovdqa %%xmm9,0x230(%%rsp)\n" );
1033     output( "\tmovdqa %%xmm10,0x240(%%rsp)\n" );
1034     output( "\tmovdqa %%xmm11,0x250(%%rsp)\n" );
1035     output( "\tmovdqa %%xmm12,0x260(%%rsp)\n" );
1036     output( "\tmovdqa %%xmm13,0x270(%%rsp)\n" );
1037     output( "\tmovdqa %%xmm14,0x280(%%rsp)\n" );
1038     output( "\tmovdqa %%xmm15,0x290(%%rsp)\n" );
1039
1040     output( "\tmovw %%cs,0x38(%%rsp)\n" );
1041     output( "\tmovw %%ds,0x3a(%%rsp)\n" );
1042     output( "\tmovw %%es,0x3c(%%rsp)\n" );
1043     output( "\tmovw %%fs,0x3e(%%rsp)\n" );
1044     output( "\tmovw %%gs,0x40(%%rsp)\n" );
1045     output( "\tmovw %%ss,0x42(%%rsp)\n" );
1046     output( "\tpushfq\n" );
1047     output( "\tpopq %%rax\n" );
1048     output( "\tmovl %%eax,0x44(%%rsp)\n" );
1049
1050     output( "\tmovl $0x%x,0x30(%%rsp)\n", 0x0010000f );
1051
1052     /* transfer the arguments */
1053
1054     output( "\tmovq %%r8,%u(%%rsp)\n", STACK_SPACE + 32 );
1055     output( "\tmovq %%r9,%u(%%rsp)\n", STACK_SPACE + 40 );
1056     output( "\tmovq $4,%%rax\n" );
1057     output( "\tleaq %u(%%rsp),%%rsi\n", STACK_SPACE + 16 );
1058     output( "\tcmpq %%rax,%%rcx\n" );
1059     output( "\tcmovgq %%rcx,%%rax\n" );
1060     output( "\tmovq %%rsp,%%rbx\n" );
1061     output( "\tleaq 16(,%%rax,8),%%rax\n" );  /* add 8 for context arg and 8 for rounding */
1062     output( "\tandq $~15,%%rax\n" );
1063     output( "\tsubq %%rax,%%rsp\n" );
1064     output( "\tmovq %%rsp,%%rdi\n" );
1065     output( "\tjrcxz 1f\n" );
1066     output( "\tcld\n" );
1067     output( "\trep\n\tmovsq\n" );
1068     output( "1:\tmovq %%rbx,0(%%rdi)\n" );  /* context arg */
1069
1070     /* call the entry point */
1071
1072     output( "\tmovq %%rdx,%%rax\n" );
1073     output( "\tmovq 0(%%rsp),%%rcx\n" );
1074     output( "\tmovq 8(%%rsp),%%rdx\n" );
1075     output( "\tmovq 16(%%rsp),%%r8\n" );
1076     output( "\tmovq 24(%%rsp),%%r9\n" );
1077     output( "\tcallq *%%rax\n" );
1078
1079     /* restore the context structure */
1080
1081     output( "1:\tmovq 0x80(%%rbx),%%rcx\n" );
1082     output( "\tmovq 0x88(%%rbx),%%rdx\n" );
1083     output( "\tmovq 0xa0(%%rbx),%%rbp\n" );
1084     output( "\tmovq 0xa8(%%rbx),%%rsi\n" );
1085     output( "\tmovq 0xb0(%%rbx),%%rdi\n" );
1086     output( "\tmovq 0xb8(%%rbx),%%r8\n" );
1087     output( "\tmovq 0xc0(%%rbx),%%r9\n" );
1088     output( "\tmovq 0xc8(%%rbx),%%r10\n" );
1089     output( "\tmovq 0xd0(%%rbx),%%r11\n" );
1090     output( "\tmovq 0xd8(%%rbx),%%r12\n" );
1091     output( "\tmovq 0xe0(%%rbx),%%r13\n" );
1092     output( "\tmovq 0xe8(%%rbx),%%r14\n" );
1093     output( "\tmovq 0xf0(%%rbx),%%r15\n" );
1094
1095     output( "\tmovdqa 0x1a0(%%rbx),%%xmm0\n" );
1096     output( "\tmovdqa 0x1b0(%%rbx),%%xmm1\n" );
1097     output( "\tmovdqa 0x1c0(%%rbx),%%xmm2\n" );
1098     output( "\tmovdqa 0x1d0(%%rbx),%%xmm3\n" );
1099     output( "\tmovdqa 0x1e0(%%rbx),%%xmm4\n" );
1100     output( "\tmovdqa 0x1f0(%%rbx),%%xmm5\n" );
1101     output( "\tmovdqa 0x200(%%rbx),%%xmm6\n" );
1102     output( "\tmovdqa 0x210(%%rbx),%%xmm7\n" );
1103     output( "\tmovdqa 0x220(%%rbx),%%xmm8\n" );
1104     output( "\tmovdqa 0x230(%%rbx),%%xmm9\n" );
1105     output( "\tmovdqa 0x240(%%rbx),%%xmm10\n" );
1106     output( "\tmovdqa 0x250(%%rbx),%%xmm11\n" );
1107     output( "\tmovdqa 0x260(%%rbx),%%xmm12\n" );
1108     output( "\tmovdqa 0x270(%%rbx),%%xmm13\n" );
1109     output( "\tmovdqa 0x280(%%rbx),%%xmm14\n" );
1110     output( "\tmovdqa 0x290(%%rbx),%%xmm15\n" );
1111     output( "\tfxrstor 0x100(%%rbx)\n" );
1112     output( "\tldmxcsr 0x34(%%rbx)\n" );
1113
1114     output( "\tmovl 0x44(%%rbx),%%eax\n" );
1115     output( "\tpushq %%rax\n" );
1116     output( "\tpopfq\n" );
1117
1118     output( "\tmovq 0x98(%%rbx),%%rax\n" );  /* stack pointer */
1119     output( "\tpushq 0xf8(%%rbx)\n" );  /* return address */
1120     output( "\tpopq -8(%%rax)\n" );
1121     output( "\tpushq 0x78(%%rbx)\n" );  /* rax */
1122     output( "\tpopq -16(%%rax)\n" );
1123     output( "\tmovq 0x90(%%rbx),%%rbx\n" );
1124     output( "\tleaq -16(%%rax),%%rsp\n" );
1125     output( "\tpopq %%rax\n" );
1126     output( "\tret\n" );
1127
1128     output_function_size( "__wine_call_from_regs" );
1129
1130     function_header( "__wine_restore_regs" );
1131     output( "\tmovq %%rcx,%%rbx\n" );
1132     output( "\tjmp 1b\n" );
1133     output_function_size( "__wine_restore_regs" );
1134 }
1135
1136
1137 /*******************************************************************
1138  *         BuildRelays32
1139  *
1140  * Build all the 32-bit relay callbacks
1141  */
1142 void BuildRelays32(void)
1143 {
1144     switch (target_cpu)
1145     {
1146     case CPU_x86:
1147         output( "/* File generated automatically. Do not edit! */\n\n" );
1148         output( "\t.text\n" );
1149         output( "%s:\n\n", asm_name("__wine_spec_thunk_text_32") );
1150
1151         /* 32-bit register entry point */
1152         BuildCallFrom32Regs();
1153
1154         output_function_size( "__wine_spec_thunk_text_32" );
1155         output_gnu_stack_note();
1156         break;
1157
1158     case CPU_x86_64:
1159         output( "/* File generated automatically. Do not edit! */\n\n" );
1160         output( "\t.text\n" );
1161         build_call_from_regs_x86_64();
1162         output_gnu_stack_note();
1163         break;
1164
1165     default:
1166         output( "/* File not used with this architecture. Do not edit! */\n\n" );
1167         return;
1168     }
1169 }