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
  11 #include <ctype.h>
  12 #include <unistd.h>
  13
  14 #include "winnt.h"
  15 #include "thread.h"
  16 #include "stackframe.h"
  17
  18 #include "build.h"
  19
  20
  21 /*******************************************************************
  22  *         BuildCallFrom16Core
  23  *
  24  * This routine builds the core routines used in 16->32 thunks:
  25  * CallFrom16Word, CallFrom16Long, CallFrom16Register, and CallFrom16Thunk.
  26  *
  27  * These routines are intended to be called via a far call (with 32-bit
  28  * operand size) from 16-bit code.  The 16-bit code stub must push %bp,
  29  * the 32-bit entry point to be called, and the argument conversion
  30  * routine to be used (see stack layout below).
  31  *
  32  * The core routine completes the STACK16FRAME on the 16-bit stack and
  33  * switches to the 32-bit stack.  Then, the argument conversion routine
  34  * is called; it gets passed the 32-bit entry point and a pointer to the
  35  * 16-bit arguments (on the 16-bit stack) as parameters. (You can either
  36  * use conversion routines automatically generated by BuildCallFrom16,
  37  * or write your own for special purposes.)
  38  *
  39  * The conversion routine must call the 32-bit entry point, passing it
  40  * the converted arguments, and return its return value to the core.
  41  * After the conversion routine has returned, the core switches back
  42  * to the 16-bit stack, converts the return value to the DX:AX format
  43  * (CallFrom16Long), and returns to the 16-bit call stub.  All parameters,
  44  * including %bp, are popped off the stack.
  45  *
  46  * The 16-bit call stub now returns to the caller, popping the 16-bit
  47  * arguments if necessary (pascal calling convention).
  48  *
  49  * In the case of a 'register' function, CallFrom16Register fills a
  50  * CONTEXT86 structure with the values all registers had at the point
  51  * the first instruction of the 16-bit call stub was about to be
  52  * executed.  A pointer to this CONTEXT86 is passed as third parameter
  53  * to the argument conversion routine, which typically passes it on
  54  * to the called 32-bit entry point.
  55  *
  56  * CallFrom16Thunk is a special variant used by the implementation of
  57  * the Win95 16->32 thunk functions C16ThkSL and C16ThkSL01 and is
  58  * implemented as follows:
  59  * On entry, the EBX register is set up to contain a flat pointer to the
  60  * 16-bit stack such that EBX+22 points to the first argument.
  61  * Then, the entry point is called, while EBP is set up to point
  62  * to the return address (on the 32-bit stack).
  63  * The called function returns with CX set to the number of bytes
  64  * to be popped of the caller's stack.
  65  *
  66  * Stack layout upon entry to the core routine (STACK16FRAME):
  67  *  ...           ...
  68  * (sp+24) word   first 16-bit arg
  69  * (sp+22) word   cs
  70  * (sp+20) word   ip
  71  * (sp+18) word   bp
  72  * (sp+14) long   32-bit entry point (reused for Win16 mutex recursion count)
  73  * (sp+12) word   ip of actual entry point (necessary for relay debugging)
  74  * (sp+8)  long   relay (argument conversion) function entry point
  75  * (sp+4)  long   cs of 16-bit entry point
  76  * (sp)    long   ip of 16-bit entry point
  77  *
  78  * Added on the stack:
  79  * (sp-2)  word   saved gs
  80  * (sp-4)  word   saved fs
  81  * (sp-6)  word   saved es
  82  * (sp-8)  word   saved ds
  83  * (sp-12) long   saved ebp
  84  * (sp-16) long   saved ecx
  85  * (sp-20) long   saved edx
  86  * (sp-24) long   saved previous stack
  87  */
  88 static void BuildCallFrom16Core( FILE *outfile, int reg_func, int thunk, int short_ret )
  89 {
  90     char *name = thunk? "Thunk" : reg_func? "Register" : short_ret? "Word" : "Long";
  91
  92     /* Function header */
  93     fprintf( outfile, "\n\t.align 4\n" );
  94 #ifdef USE_STABS
  95     fprintf( outfile, ".stabs \"CallFrom16%s:F1\",36,0,0," PREFIX "CallFrom16%s\n",
  96              name, name);
  97 #endif
  98     fprintf( outfile, "\t.type " PREFIX "CallFrom16%s,@function\n", name );
  99     fprintf( outfile, "\t.globl " PREFIX "CallFrom16%s\n", name );
 100     fprintf( outfile, PREFIX "CallFrom16%s:\n", name );
 101
 102     /* Create STACK16FRAME (except STACK32FRAME link) */
 103     fprintf( outfile, "\tpushw %%gs\n" );
 104     fprintf( outfile, "\tpushw %%fs\n" );
 105     fprintf( outfile, "\tpushw %%es\n" );
 106     fprintf( outfile, "\tpushw %%ds\n" );
 107     fprintf( outfile, "\tpushl %%ebp\n" );
 108     fprintf( outfile, "\tpushl %%ecx\n" );
 109     fprintf( outfile, "\tpushl %%edx\n" );
 110
 111     /* Save original EFlags register */
 112     fprintf( outfile, "\tpushfl\n" );
 113
 114     if ( UsePIC )
 115     {
 116         /* Get Global Offset Table into %ecx */
 117         fprintf( outfile, "\tcall .LCallFrom16%s.getgot1\n", name );
 118         fprintf( outfile, ".LCallFrom16%s.getgot1:\n", name );
 119         fprintf( outfile, "\tpopl %%ecx\n" );
 120         fprintf( outfile, "\taddl $_GLOBAL_OFFSET_TABLE_+[.-.LCallFrom16%s.getgot1], %%ecx\n", name );
 121     }
 122
 123     /* Load 32-bit segment registers */
 124     fprintf( outfile, "\tmovw $0x%04x, %%dx\n", data_selector );
 125 #ifdef __svr4__
 126     fprintf( outfile, "\tdata16\n");
 127 #endif
 128     fprintf( outfile, "\tmovw %%dx, %%ds\n" );
 129 #ifdef __svr4__
 130     fprintf( outfile, "\tdata16\n");
 131 #endif
 132     fprintf( outfile, "\tmovw %%dx, %%es\n" );
 133
 134     if ( UsePIC )
 135     {
 136         fprintf( outfile, "\tmovl " PREFIX "SYSLEVEL_Win16CurrentTeb@GOT(%%ecx), %%edx\n" );
 137         fprintf( outfile, "\tmovw (%%edx), %%fs\n" );
 138     }
 139     else
 140         fprintf( outfile, "\tmovw " PREFIX "SYSLEVEL_Win16CurrentTeb, %%fs\n" );
 141
 142     /* Get address of ldt_copy array into %ecx */
 143     if ( UsePIC )
 144         fprintf( outfile, "\tmovl " PREFIX "ldt_copy@GOT(%%ecx), %%ecx\n" );
 145     else
 146         fprintf( outfile, "\tmovl $" PREFIX "ldt_copy, %%ecx\n" );
 147
 148     /* Translate STACK16FRAME base to flat offset in %edx */
 149     fprintf( outfile, "\tmovw %%ss, %%dx\n" );
 150     fprintf( outfile, "\tandl $0xfff8, %%edx\n" );
 151     fprintf( outfile, "\tmovl (%%ecx,%%edx), %%edx\n" );
 152     fprintf( outfile, "\tmovzwl %%sp, %%ebp\n" );
 153     fprintf( outfile, "\tleal (%%ebp,%%edx), %%edx\n" );
 154
 155     /* Get saved flags into %ecx */
 156     fprintf( outfile, "\tpopl %%ecx\n" );
 157
 158     /* Get the 32-bit stack pointer from the TEB and complete STACK16FRAME */
 159     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d), %%ebp\n", STACKOFFSET );
 160     fprintf( outfile, "\tpushl %%ebp\n" );
 161
 162     /* Switch stacks */
 163 #ifdef __svr4__
 164     fprintf( outfile,"\tdata16\n");
 165 #endif
 166     fprintf( outfile, "\t.byte 0x64\n\tmovw %%ss, (%d)\n", STACKOFFSET + 2 );
 167     fprintf( outfile, "\t.byte 0x64\n\tmovw %%sp, (%d)\n", STACKOFFSET );
 168     fprintf( outfile, "\tpushl %%ds\n" );
 169     fprintf( outfile, "\tpopl %%ss\n" );
 170     fprintf( outfile, "\tmovl %%ebp, %%esp\n" );
 171     fprintf( outfile, "\taddl $%d, %%ebp\n", STRUCTOFFSET(STACK32FRAME, ebp) );
 172
 173
 174     /* At this point:
 175        STACK16FRAME is completely set up
 176        DS, ES, SS: flat data segment
 177        FS: current TEB
 178        ESP: points to last STACK32FRAME
 179        EBP: points to ebp member of last STACK32FRAME
 180        EDX: points to current STACK16FRAME
 181        ECX: contains saved flags
 182        all other registers: unchanged */
 183
 184     /* Special case: C16ThkSL stub */
 185     if ( thunk )
 186     {
 187         /* Set up registers as expected and call thunk */
 188         fprintf( outfile, "\tleal %d(%%edx), %%ebx\n", sizeof(STACK16FRAME)-22 );
 189         fprintf( outfile, "\tleal -4(%%esp), %%ebp\n" );
 190
 191         fprintf( outfile, "\tcall *%d(%%edx)\n", STACK16OFFSET(entry_point) );
 192
 193         /* Switch stack back */
 194         /* fprintf( outfile, "\t.byte 0x64\n\tlssw (%d), %%sp\n", STACKOFFSET ); */
 195         fprintf( outfile, "\t.byte 0x64,0x66,0x0f,0xb2,0x25\n\t.long %d\n", STACKOFFSET );
 196         fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 197
 198         /* Restore registers and return directly to caller */
 199         fprintf( outfile, "\taddl $8, %%esp\n" );
 200         fprintf( outfile, "\tpopl %%ebp\n" );
 201         fprintf( outfile, "\tpopw %%ds\n" );
 202         fprintf( outfile, "\tpopw %%es\n" );
 203         fprintf( outfile, "\tpopw %%fs\n" );
 204         fprintf( outfile, "\tpopw %%gs\n" );
 205         fprintf( outfile, "\taddl $20, %%esp\n" );
 206
 207         fprintf( outfile, "\txorb %%ch, %%ch\n" );
 208         fprintf( outfile, "\tpopl %%ebx\n" );
 209         fprintf( outfile, "\taddw %%cx, %%sp\n" );
 210         fprintf( outfile, "\tpush %%ebx\n" );
 211
 212         fprintf( outfile, "\t.byte 0x66\n" );
 213         fprintf( outfile, "\tlret\n" );
 214
 215         return;
 216     }
 217
 218
 219     /* Build register CONTEXT */
 220     if ( reg_func )
 221     {
 222         fprintf( outfile, "\tsubl $%d, %%esp\n", sizeof(CONTEXT86) );
 223
 224         fprintf( outfile, "\tmovl %%ecx, %d(%%esp)\n", CONTEXTOFFSET(EFlags) );
 225
 226         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Eax) );
 227         fprintf( outfile, "\tmovl %%ebx, %d(%%esp)\n", CONTEXTOFFSET(Ebx) );
 228         fprintf( outfile, "\tmovl %%esi, %d(%%esp)\n", CONTEXTOFFSET(Esi) );
 229         fprintf( outfile, "\tmovl %%edi, %d(%%esp)\n", CONTEXTOFFSET(Edi) );
 230
 231         fprintf( outfile, "\tmovl %d(%%edx), %%eax\n", STACK16OFFSET(ebp) );
 232         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Ebp) );
 233         fprintf( outfile, "\tmovl %d(%%edx), %%eax\n", STACK16OFFSET(ecx) );
 234         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Ecx) );
 235         fprintf( outfile, "\tmovl %d(%%edx), %%eax\n", STACK16OFFSET(edx) );
 236         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Edx) );
 237
 238         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(ds) );
 239         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegDs) );
 240         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(es) );
 241         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegEs) );
 242         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(fs) );
 243         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegFs) );
 244         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(gs) );
 245         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegGs) );
 246
 247         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(cs) );
 248         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegCs) );
 249         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(ip) );
 250         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Eip) );
 251
 252         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%eax\n", STACKOFFSET+2 );
 253         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegSs) );
 254         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%eax\n", STACKOFFSET );
 255         fprintf( outfile, "\taddl $%d, %%eax\n", STACK16OFFSET(ip) );
 256         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Esp) );
 257 #if 0
 258         fprintf( outfile, "\tfsave %d(%%esp)\n", CONTEXTOFFSET(FloatSave) );
 259 #endif
 260
 261         /* Push address of CONTEXT86 structure -- popped by the relay routine */
 262         fprintf( outfile, "\tpushl %%esp\n" );
 263     }
 264
 265
 266     /* Print debug info before call */
 267     if ( debugging )
 268     {
 269         if ( UsePIC )
 270         {
 271             fprintf( outfile, "\tpushl %%ebx\n" );
 272
 273             /* Get Global Offset Table into %ebx (for PLT call) */
 274             fprintf( outfile, "\tcall .LCallFrom16%s.getgot2\n", name );
 275             fprintf( outfile, ".LCallFrom16%s.getgot2:\n", name );
 276             fprintf( outfile, "\tpopl %%ebx\n" );
 277             fprintf( outfile, "\taddl $_GLOBAL_OFFSET_TABLE_+[.-.LCallFrom16%s.getgot2], %%ebx\n", name );
 278         }
 279
 280         fprintf( outfile, "\tpushl %%edx\n" );
 281         if ( reg_func )
 282             fprintf( outfile, "\tleal -%d(%%ebp), %%eax\n\tpushl %%eax\n",
 283                               sizeof(CONTEXT) + STRUCTOFFSET(STACK32FRAME, ebp) );
 284         else
 285             fprintf( outfile, "\tpushl $0\n" );
 286
 287         if ( UsePIC )
 288             fprintf( outfile, "\tcall " PREFIX "RELAY_DebugCallFrom16@PLT\n ");
 289         else
 290             fprintf( outfile, "\tcall " PREFIX "RELAY_DebugCallFrom16\n ");
 291
 292         fprintf( outfile, "\tpopl %%edx\n" );
 293         fprintf( outfile, "\tpopl %%edx\n" );
 294
 295         if ( UsePIC )
 296             fprintf( outfile, "\tpopl %%ebx\n" );
 297     }
 298
 299     /* Call relay routine (which will call the API entry point) */
 300     fprintf( outfile, "\tleal %d(%%edx), %%eax\n", sizeof(STACK16FRAME) );
 301     fprintf( outfile, "\tpushl %%eax\n" );
 302     fprintf( outfile, "\tpushl %d(%%edx)\n", STACK16OFFSET(entry_point) );
 303     fprintf( outfile, "\tcall *%d(%%edx)\n", STACK16OFFSET(relay) );
 304
 305     /* Print debug info after call */
 306     if ( debugging )
 307     {
 308         if ( UsePIC )
 309         {
 310             fprintf( outfile, "\tpushl %%ebx\n" );
 311
 312             /* Get Global Offset Table into %ebx (for PLT call) */
 313             fprintf( outfile, "\tcall .LCallFrom16%s.getgot3\n", name );
 314             fprintf( outfile, ".LCallFrom16%s.getgot3:\n", name );
 315             fprintf( outfile, "\tpopl %%ebx\n" );
 316             fprintf( outfile, "\taddl $_GLOBAL_OFFSET_TABLE_+[.-.LCallFrom16%s.getgot3], %%ebx\n", name );
 317         }
 318
 319         fprintf( outfile, "\tpushl %%eax\n" );
 320         if ( reg_func )
 321             fprintf( outfile, "\tleal -%d(%%ebp), %%eax\n\tpushl %%eax\n",
 322                               sizeof(CONTEXT) + STRUCTOFFSET(STACK32FRAME, ebp) );
 323         else
 324             fprintf( outfile, "\tpushl $0\n" );
 325
 326         if ( UsePIC )
 327             fprintf( outfile, "\tcall " PREFIX "RELAY_DebugCallFrom16Ret@PLT\n ");
 328         else
 329             fprintf( outfile, "\tcall " PREFIX "RELAY_DebugCallFrom16Ret\n ");
 330
 331         fprintf( outfile, "\tpopl %%eax\n" );
 332         fprintf( outfile, "\tpopl %%eax\n" );
 333
 334         if ( UsePIC )
 335             fprintf( outfile, "\tpopl %%ebx\n" );
 336     }
 337
 338
 339     if ( reg_func )
 340     {
 341         fprintf( outfile, "\tmovl %%esp, %%ebx\n" );
 342
 343         /* Switch stack back */
 344         /* fprintf( outfile, "\t.byte 0x64\n\tlssw (%d), %%sp\n", STACKOFFSET ); */
 345         fprintf( outfile, "\t.byte 0x64,0x66,0x0f,0xb2,0x25\n\t.long %d\n", STACKOFFSET );
 346         fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 347
 348         /* Get return address to CallFrom16 stub */
 349         fprintf( outfile, "\taddw $%d, %%sp\n", STACK16OFFSET(callfrom_ip)-4 );
 350         fprintf( outfile, "\tpopl %%eax\n" );
 351         fprintf( outfile, "\tpopl %%edx\n" );
 352
 353         /* Restore all registers from CONTEXT */
 354         fprintf( outfile, "\tmovw %d(%%ebx), %%ss\n", CONTEXTOFFSET(SegSs) );
 355         fprintf( outfile, "\tmovl %d(%%ebx), %%esp\n", CONTEXTOFFSET(Esp) );
 356         fprintf( outfile, "\taddl $4, %%esp\n" );  /* room for final return address */
 357
 358         fprintf( outfile, "\tpushw %d(%%ebx)\n", CONTEXTOFFSET(SegCs) );
 359         fprintf( outfile, "\tpushw %d(%%ebx)\n", CONTEXTOFFSET(Eip) );
 360         fprintf( outfile, "\tpushl %%edx\n" );
 361         fprintf( outfile, "\tpushl %%eax\n" );
 362         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(EFlags) );
 363         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(SegDs) );
 364
 365         fprintf( outfile, "\tmovw %d(%%ebx), %%es\n", CONTEXTOFFSET(SegEs) );
 366         fprintf( outfile, "\tmovw %d(%%ebx), %%fs\n", CONTEXTOFFSET(SegFs) );
 367         fprintf( outfile, "\tmovw %d(%%ebx), %%gs\n", CONTEXTOFFSET(SegGs) );
 368
 369         fprintf( outfile, "\tmovl %d(%%ebx), %%ebp\n", CONTEXTOFFSET(Ebp) );
 370         fprintf( outfile, "\tmovl %d(%%ebx), %%esi\n", CONTEXTOFFSET(Esi) );
 371         fprintf( outfile, "\tmovl %d(%%ebx), %%edi\n", CONTEXTOFFSET(Edi) );
 372         fprintf( outfile, "\tmovl %d(%%ebx), %%eax\n", CONTEXTOFFSET(Eax) );
 373         fprintf( outfile, "\tmovl %d(%%ebx), %%edx\n", CONTEXTOFFSET(Edx) );
 374         fprintf( outfile, "\tmovl %d(%%ebx), %%ecx\n", CONTEXTOFFSET(Ecx) );
 375         fprintf( outfile, "\tmovl %d(%%ebx), %%ebx\n", CONTEXTOFFSET(Ebx) );
 376
 377         fprintf( outfile, "\tpopl %%ds\n" );
 378         fprintf( outfile, "\tpopfl\n" );
 379         fprintf( outfile, "\tlret\n" );
 380     }
 381     else
 382     {
 383         /* Switch stack back */
 384         /* fprintf( outfile, "\t.byte 0x64\n\tlssw (%d), %%sp\n", STACKOFFSET ); */
 385         fprintf( outfile, "\t.byte 0x64,0x66,0x0f,0xb2,0x25\n\t.long %d\n", STACKOFFSET );
 386         fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 387
 388         /* Restore registers */
 389         fprintf( outfile, "\tpopl %%edx\n" );
 390         fprintf( outfile, "\tpopl %%ecx\n" );
 391         fprintf( outfile, "\tpopl %%ebp\n" );
 392         fprintf( outfile, "\tpopw %%ds\n" );
 393         fprintf( outfile, "\tpopw %%es\n" );
 394         fprintf( outfile, "\tpopw %%fs\n" );
 395         fprintf( outfile, "\tpopw %%gs\n" );
 396
 397         /* Prepare return value and set flags accordingly */
 398         if ( !short_ret )
 399             fprintf( outfile, "\tshldl $16, %%eax, %%edx\n" );
 400         fprintf( outfile, "\torl %%eax, %%eax\n" );
 401
 402         /* Return to return stub which will return to caller */
 403         fprintf( outfile, "\tlret $12\n" );
 404     }
 405 }
 406
 407
 408 /*******************************************************************
 409  *         BuildCallTo16Core
 410  *
 411  * This routine builds the core routines used in 32->16 thunks:
 412  *
 413  *   extern void CALLBACK CallTo16Word( SEGPTR target, int nb_args );
 414  *   extern void CALLBACK CallTo16Long( SEGPTR target, int nb_args );
 415  *   extern void CALLBACK CallTo16RegisterShort( const CONTEXT86 *context, int nb_args );
 416  *   extern void CALLBACK CallTo16RegisterLong ( const CONTEXT86 *context, int nb_args );
 417  *
 418  * These routines can be called directly from 32-bit code.
 419  *
 420  * All routines expect that the 16-bit stack contents (arguments) were
 421  * already set up by the caller; nb_args must contain the number of bytes
 422  * to be conserved.  The 16-bit SS:SP will be set accordinly.
 423  *
 424  * All other registers are either taken from the CONTEXT86 structure
 425  * or else set to default values.  The target routine address is either
 426  * given directly or taken from the CONTEXT86.
 427  *
 428  * If you want to call a 16-bit routine taking only standard argument types
 429  * (WORD and LONG), you can also have an appropriate argument conversion
 430  * stub automatically generated (see BuildCallTo16); you'd then call this
 431  * stub, which in turn would prepare the 16-bit stack and call the appropiate
 432  * core routine.
 433  *
 434  */
 435
 436 static void BuildCallTo16Core( FILE *outfile, int short_ret, int reg_func )
 437 {
 438     char *name = reg_func == 2 ? "RegisterLong" :
 439                  reg_func == 1 ? "RegisterShort" :
 440                  short_ret? "Word" : "Long";
 441
 442     /* Function header */
 443     fprintf( outfile, "\n\t.align 4\n" );
 444 #ifdef USE_STABS
 445     fprintf( outfile, ".stabs \"CallTo16%s:F1\",36,0,0," PREFIX "CallTo16%s\n",
 446              name, name);
 447 #endif
 448     fprintf( outfile, "\t.type " PREFIX "CallTo16%s,@function\n", name );
 449     fprintf( outfile, "\t.globl " PREFIX "CallTo16%s\n", name );
 450     fprintf( outfile, PREFIX "CallTo16%s:\n", name );
 451
 452     /* Function entry sequence */
 453     fprintf( outfile, "\tpushl %%ebp\n" );
 454     fprintf( outfile, "\tmovl %%esp, %%ebp\n" );
 455
 456     /* Save the 32-bit registers */
 457     fprintf( outfile, "\tpushl %%ebx\n" );
 458     fprintf( outfile, "\tpushl %%ecx\n" );
 459     fprintf( outfile, "\tpushl %%edx\n" );
 460     fprintf( outfile, "\tpushl %%esi\n" );
 461     fprintf( outfile, "\tpushl %%edi\n" );
 462
 463     if ( UsePIC )
 464     {
 465         /* Get Global Offset Table into %ebx */
 466         fprintf( outfile, "\tcall .LCallTo16%s.getgot1\n", name );
 467         fprintf( outfile, ".LCallTo16%s.getgot1:\n", name );
 468         fprintf( outfile, "\tpopl %%ebx\n" );
 469         fprintf( outfile, "\taddl $_GLOBAL_OFFSET_TABLE_+[.-.LCallTo16%s.getgot1], %%ebx\n", name );
 470     }
 471
 472     /* Enter Win16 Mutex */
 473     if ( UsePIC )
 474         fprintf( outfile, "\tcall " PREFIX "SYSLEVEL_EnterWin16Lock@PLT\n" );
 475     else
 476         fprintf( outfile, "\tcall " PREFIX "SYSLEVEL_EnterWin16Lock\n" );
 477
 478     /* Print debugging info */
 479     if (debugging)
 480     {
 481         /* Push flags, number of arguments, and target */
 482         fprintf( outfile, "\tpushl $%d\n", reg_func );
 483         fprintf( outfile, "\tpushl 12(%%ebp)\n" );
 484         fprintf( outfile, "\tpushl  8(%%ebp)\n" );
 485
 486         if ( UsePIC )
 487             fprintf( outfile, "\tcall " PREFIX "RELAY_DebugCallTo16@PLT\n" );
 488         else
 489             fprintf( outfile, "\tcall " PREFIX "RELAY_DebugCallTo16\n" );
 490
 491         fprintf( outfile, "\taddl $12, %%esp\n" );
 492     }
 493
 494     /* Get return address */
 495     if ( UsePIC )
 496     {
 497         fprintf( outfile, "\tmovl " PREFIX "CallTo16_RetAddr@GOT(%%ebx), %%ecx\n" );
 498         fprintf( outfile, "\tmovl " PREFIX "(%%ecx), %%ecx\n" );
 499     }
 500     else
 501         fprintf( outfile, "\tmovl " PREFIX "CallTo16_RetAddr, %%ecx\n" );
 502
 503     /* Call the actual CallTo16 routine (simulate a lcall) */
 504     fprintf( outfile, "\tpushl %%cs\n" );
 505     fprintf( outfile, "\tcall .LCallTo16%s\n", name );
 506
 507     /* Convert and push return value */
 508     if ( short_ret )
 509     {
 510         fprintf( outfile, "\tmovzwl %%ax, %%eax\n" );
 511         fprintf( outfile, "\tpushl %%eax\n" );
 512     }
 513     else if ( reg_func != 2 )
 514     {
 515         fprintf( outfile, "\tshll $16,%%edx\n" );
 516         fprintf( outfile, "\tmovw %%ax,%%dx\n" );
 517         fprintf( outfile, "\tpushl %%edx\n" );
 518     }
 519     else
 520         fprintf( outfile, "\tpushl %%eax\n" );
 521
 522     if ( UsePIC )
 523     {
 524         /* Get Global Offset Table into %ebx (might have been overwritten) */
 525         fprintf( outfile, "\tcall .LCallTo16%s.getgot2\n", name );
 526         fprintf( outfile, ".LCallTo16%s.getgot2:\n", name );
 527         fprintf( outfile, "\tpopl %%ebx\n" );
 528         fprintf( outfile, "\taddl $_GLOBAL_OFFSET_TABLE_+[.-.LCallTo16%s.getgot2], %%ebx\n", name );
 529     }
 530
 531     /* Print debugging info */
 532     if (debugging)
 533     {
 534         if ( UsePIC )
 535             fprintf( outfile, "\tcall " PREFIX "RELAY_DebugCallTo16Ret@PLT\n" );
 536         else
 537             fprintf( outfile, "\tcall " PREFIX "RELAY_DebugCallTo16Ret\n" );
 538     }
 539
 540     /* Leave Win16 Mutex */
 541     if ( UsePIC )
 542         fprintf( outfile, "\tcall " PREFIX "SYSLEVEL_LeaveWin16Lock@PLT\n" );
 543     else
 544         fprintf( outfile, "\tcall " PREFIX "SYSLEVEL_LeaveWin16Lock\n" );
 545
 546     /* Get return value */
 547     fprintf( outfile, "\tpopl %%eax\n" );
 548
 549     /* Restore the 32-bit registers */
 550     fprintf( outfile, "\tpopl %%edi\n" );
 551     fprintf( outfile, "\tpopl %%esi\n" );
 552     fprintf( outfile, "\tpopl %%edx\n" );
 553     fprintf( outfile, "\tpopl %%ecx\n" );
 554     fprintf( outfile, "\tpopl %%ebx\n" );
 555
 556     /* Function exit sequence */
 557     fprintf( outfile, "\tpopl %%ebp\n" );
 558     fprintf( outfile, "\tret $8\n" );
 559
 560
 561     /* Start of the actual CallTo16 routine */
 562
 563     fprintf( outfile, ".LCallTo16%s:\n", name );
 564
 565     /* Complete STACK32FRAME */
 566     fprintf( outfile, "\t.byte 0x64\n\tpushl (%d)\n", STACKOFFSET );
 567     fprintf( outfile, "\tmovl %%esp,%%edx\n" );
 568
 569     /* Switch to the 16-bit stack */
 570 #ifdef __svr4__
 571     fprintf( outfile,"\tdata16\n");
 572 #endif
 573     fprintf( outfile, "\t.byte 0x64\n\tmovw (%d),%%ss\n", STACKOFFSET + 2);
 574     fprintf( outfile, "\t.byte 0x64\n\tmovw (%d),%%sp\n", STACKOFFSET );
 575     fprintf( outfile, "\t.byte 0x64\n\tmovl %%edx,(%d)\n", STACKOFFSET );
 576
 577     /* Make %bp point to the previous stackframe (built by CallFrom16) */
 578     fprintf( outfile, "\tmovzwl %%sp,%%ebp\n" );
 579     fprintf( outfile, "\tleal %d(%%ebp),%%ebp\n", STACK16OFFSET(bp) );
 580
 581     /* Add the specified offset to the new sp */
 582     fprintf( outfile, "\tsubw %d(%%edx), %%sp\n", STACK32OFFSET(nb_args) );
 583
 584     /* Push the return address
 585      * With sreg suffix, we push 16:16 address (normal lret)
 586      * With lreg suffix, we push 16:32 address (0x66 lret, for KERNEL32_45)
 587      */
 588     if (reg_func != 2)
 589         fprintf( outfile, "\tpushl %%ecx\n" );
 590     else
 591     {
 592         fprintf( outfile, "\tshldl $16, %%ecx, %%eax\n" );
 593         fprintf( outfile, "\tpushw $0\n" );
 594         fprintf( outfile, "\tpushw %%ax\n" );
 595         fprintf( outfile, "\tpushw $0\n" );
 596         fprintf( outfile, "\tpushw %%cx\n" );
 597     }
 598
 599     if (reg_func)
 600     {
 601         /* Push the called routine address */
 602         fprintf( outfile, "\tmovl %d(%%edx),%%edx\n", STACK32OFFSET(target) );
 603         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(SegCs) );
 604         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(Eip) );
 605
 606         /* Get the registers */
 607         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(SegDs) );
 608         fprintf( outfile, "\tmovl %d(%%edx),%%eax\n", CONTEXTOFFSET(SegEs) );
 609         fprintf( outfile, "\tmovw %%ax,%%es\n" );
 610         fprintf( outfile, "\tmovl %d(%%edx),%%eax\n", CONTEXTOFFSET(SegFs) );
 611         fprintf( outfile, "\tmovw %%ax,%%fs\n" );
 612         fprintf( outfile, "\tmovl %d(%%edx),%%ebp\n", CONTEXTOFFSET(Ebp) );
 613         fprintf( outfile, "\tmovl %d(%%edx),%%esi\n", CONTEXTOFFSET(Esi) );
 614         fprintf( outfile, "\tmovl %d(%%edx),%%edi\n", CONTEXTOFFSET(Edi) );
 615         fprintf( outfile, "\tmovl %d(%%edx),%%eax\n", CONTEXTOFFSET(Eax) );
 616         fprintf( outfile, "\tmovl %d(%%edx),%%ebx\n", CONTEXTOFFSET(Ebx) );
 617         fprintf( outfile, "\tmovl %d(%%edx),%%ecx\n", CONTEXTOFFSET(Ecx) );
 618         fprintf( outfile, "\tmovl %d(%%edx),%%edx\n", CONTEXTOFFSET(Edx) );
 619
 620         /* Get the 16-bit ds */
 621         fprintf( outfile, "\tpopw %%ds\n" );
 622     }
 623     else  /* not a register function */
 624     {
 625         /* Push the called routine address */
 626         fprintf( outfile, "\tpushl %d(%%edx)\n", STACK32OFFSET(target) );
 627
 628         /* Set %fs to the value saved by the last CallFrom16 */
 629         fprintf( outfile, "\tmovw %d(%%ebp),%%ax\n", STACK16OFFSET(fs)-STACK16OFFSET(bp) );
 630         fprintf( outfile, "\tmovw %%ax,%%fs\n" );
 631
 632         /* Set %ds and %es (and %ax just in case) equal to %ss */
 633         fprintf( outfile, "\tmovw %%ss,%%ax\n" );
 634         fprintf( outfile, "\tmovw %%ax,%%ds\n" );
 635         fprintf( outfile, "\tmovw %%ax,%%es\n" );
 636     }
 637
 638     /* Jump to the called routine */
 639     fprintf( outfile, "\t.byte 0x66\n" );
 640     fprintf( outfile, "\tlret\n" );
 641 }
 642
 643
 644 /*******************************************************************
 645  *         BuildRet16Func
 646  *
 647  * Build the return code for 16-bit callbacks
 648  */
 649 static void BuildRet16Func( FILE *outfile )
 650 {
 651     /*
 652      *  Note: This must reside in the .data section to allow
 653      *        run-time relocation of the SYSLEVEL_Win16CurrentTeb symbol
 654      */
 655
 656     fprintf( outfile, "\n\t.type " PREFIX "CallTo16_Ret,@function\n" );
 657     fprintf( outfile, "\t.globl " PREFIX "CallTo16_Ret\n" );
 658     fprintf( outfile, PREFIX "CallTo16_Ret:\n" );
 659
 660     /* Restore 32-bit segment registers */
 661
 662     fprintf( outfile, "\tmovw $0x%04x,%%bx\n", data_selector );
 663 #ifdef __svr4__
 664     fprintf( outfile, "\tdata16\n");
 665 #endif
 666     fprintf( outfile, "\tmovw %%bx,%%ds\n" );
 667 #ifdef __svr4__
 668     fprintf( outfile, "\tdata16\n");
 669 #endif
 670     fprintf( outfile, "\tmovw %%bx,%%es\n" );
 671
 672     fprintf( outfile, "\tmovw " PREFIX "SYSLEVEL_Win16CurrentTeb,%%fs\n" );
 673
 674     /* Restore the 32-bit stack */
 675
 676 #ifdef __svr4__
 677     fprintf( outfile, "\tdata16\n");
 678 #endif
 679     fprintf( outfile, "\tmovw %%bx,%%ss\n" );
 680     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d),%%esp\n", STACKOFFSET );
 681     fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 682
 683     /* Return to caller */
 684
 685     fprintf( outfile, "\tlret\n" );
 686
 687     /* Declare the return address variable */
 688
 689     fprintf( outfile, "\n\t.globl " PREFIX "CallTo16_RetAddr\n" );
 690     fprintf( outfile, PREFIX "CallTo16_RetAddr:\t.long 0\n" );
 691 }
 692
 693
 694 /*******************************************************************
 695  *         BuildCallTo32CBClient
 696  *
 697  * Call a CBClient relay stub from 32-bit code (KERNEL.620).
 698  *
 699  * Since the relay stub is itself 32-bit, this should not be a problem;
 700  * unfortunately, the relay stubs are expected to switch back to a
 701  * 16-bit stack (and 16-bit code) after completion :-(
 702  *
 703  * This would conflict with our 16- vs. 32-bit stack handling, so
 704  * we simply switch *back* to our 32-bit stack before returning to
 705  * the caller ...
 706  *
 707  * The CBClient relay stub expects to be called with the following
 708  * 16-bit stack layout, and with ebp and ebx pointing into the 16-bit
 709  * stack at the designated places:
 710  *
 711  *    ...
 712  *  (ebp+14) original arguments to the callback routine
 713  *  (ebp+10) far return address to original caller
 714  *  (ebp+6)  Thunklet target address
 715  *  (ebp+2)  Thunklet relay ID code
 716  *  (ebp)    BP (saved by CBClientGlueSL)
 717  *  (ebp-2)  SI (saved by CBClientGlueSL)
 718  *  (ebp-4)  DI (saved by CBClientGlueSL)
 719  *  (ebp-6)  DS (saved by CBClientGlueSL)
 720  *
 721  *   ...     buffer space used by the 16-bit side glue for temp copies
 722  *
 723  *  (ebx+4)  far return address to 16-bit side glue code
 724  *  (ebx)    saved 16-bit ss:sp (pointing to ebx+4)
 725  *
 726  * The 32-bit side glue code accesses both the original arguments (via ebp)
 727  * and the temporary copies prepared by the 16-bit side glue (via ebx).
 728  * After completion, the stub will load ss:sp from the buffer at ebx
 729  * and perform a far return to 16-bit code.
 730  *
 731  * To trick the relay stub into returning to us, we replace the 16-bit
 732  * return address to the glue code by a cs:ip pair pointing to our
 733  * return entry point (the original return address is saved first).
 734  * Our return stub thus called will then reload the 32-bit ss:esp and
 735  * return to 32-bit code (by using and ss:esp value that we have also
 736  * pushed onto the 16-bit stack before and a cs:eip values found at
 737  * that position on the 32-bit stack).  The ss:esp to be restored is
 738  * found relative to the 16-bit stack pointer at:
 739  *
 740  *  (ebx-4)   ss  (flat)
 741  *  (ebx-8)   sp  (32-bit stack pointer)
 742  *
 743  * The second variant of this routine, CALL32_CBClientEx, which is used
 744  * to implement KERNEL.621, has to cope with yet another problem: Here,
 745  * the 32-bit side directly returns to the caller of the CBClient thunklet,
 746  * restoring registers saved by CBClientGlueSL and cleaning up the stack.
 747  * As we have to return to our 32-bit code first, we have to adapt the
 748  * layout of our temporary area so as to include values for the registers
 749  * that are to be restored, and later (in the implementation of KERNEL.621)
 750  * we *really* restore them. The return stub restores DS, DI, SI, and BP
 751  * from the stack, skips the next 8 bytes (CBClient relay code / target),
 752  * and then performs a lret NN, where NN is the number of arguments to be
 753  * removed. Thus, we prepare our temporary area as follows:
 754  *
 755  *     (ebx+22) 16-bit cs  (this segment)
 756  *     (ebx+20) 16-bit ip  ('16-bit' return entry point)
 757  *     (ebx+16) 32-bit ss  (flat)
 758  *     (ebx+12) 32-bit sp  (32-bit stack pointer)
 759  *     (ebx+10) 16-bit bp  (points to ebx+24)
 760  *     (ebx+8)  16-bit si  (ignored)
 761  *     (ebx+6)  16-bit di  (ignored)
 762  *     (ebx+4)  16-bit ds  (we actually use the flat DS here)
 763  *     (ebx+2)  16-bit ss  (16-bit stack segment)
 764  *     (ebx+0)  16-bit sp  (points to ebx+4)
 765  *
 766  * Note that we ensure that DS is not changed and remains the flat segment,
 767  * and the 32-bit stack pointer our own return stub needs fits just
 768  * perfectly into the 8 bytes that are skipped by the Windows stub.
 769  * One problem is that we have to determine the number of removed arguments,
 770  * as these have to be really removed in KERNEL.621. Thus, the BP value
 771  * that we place in the temporary area to be restored, contains the value
 772  * that SP would have if no arguments were removed. By comparing the actual
 773  * value of SP with this value in our return stub we can compute the number
 774  * of removed arguments. This is then returned to KERNEL.621.
 775  *
 776  * The stack layout of this function:
 777  * (ebp+20)  nArgs     pointer to variable receiving nr. of args (Ex only)
 778  * (ebp+16)  esi       pointer to caller's esi value
 779  * (ebp+12)  arg       ebp value to be set for relay stub
 780  * (ebp+8)   func      CBClient relay stub address
 781  * (ebp+4)   ret addr
 782  * (ebp)     ebp
 783  */
 784 static void BuildCallTo32CBClient( FILE *outfile, BOOL isEx )
 785 {
 786     char *name = isEx? "CBClientEx" : "CBClient";
 787     int size = isEx? 24 : 12;
 788
 789     /* Function header */
 790
 791     fprintf( outfile, "\n\t.align 4\n" );
 792 #ifdef USE_STABS
 793     fprintf( outfile, ".stabs \"CALL32_%s:F1\",36,0,0," PREFIX "CALL32_%s\n",
 794                       name, name );
 795 #endif
 796     fprintf( outfile, "\t.globl " PREFIX "CALL32_%s\n", name );
 797     fprintf( outfile, PREFIX "CALL32_%s:\n", name );
 798
 799     /* Entry code */
 800
 801     fprintf( outfile, "\tpushl %%ebp\n" );
 802     fprintf( outfile, "\tmovl %%esp,%%ebp\n" );
 803     fprintf( outfile, "\tpushl %%edi\n" );
 804     fprintf( outfile, "\tpushl %%esi\n" );
 805     fprintf( outfile, "\tpushl %%ebx\n" );
 806
 807     /* Get the 16-bit stack */
 808
 809     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d),%%ebx\n", STACKOFFSET);
 810
 811     /* Convert it to a flat address */
 812
 813     fprintf( outfile, "\tshldl $16,%%ebx,%%eax\n" );
 814     fprintf( outfile, "\tandl $0xfff8,%%eax\n" );
 815     fprintf( outfile, "\tmovl " PREFIX "ldt_copy(%%eax),%%esi\n" );
 816     fprintf( outfile, "\tmovw %%bx,%%ax\n" );
 817     fprintf( outfile, "\taddl %%eax,%%esi\n" );
 818
 819     /* Allocate temporary area (simulate STACK16_PUSH) */
 820
 821     fprintf( outfile, "\tpushf\n" );
 822     fprintf( outfile, "\tcld\n" );
 823     fprintf( outfile, "\tleal -%d(%%esi), %%edi\n", size );
 824     fprintf( outfile, "\tmovl $%d, %%ecx\n", sizeof(STACK16FRAME) );
 825     fprintf( outfile, "\trep\n\tmovsb\n" );
 826     fprintf( outfile, "\tpopf\n" );
 827
 828     fprintf( outfile, "\t.byte 0x64\n\tsubw $%d,(%d)\n", size, STACKOFFSET );
 829
 830     fprintf( outfile, "\tpushl %%edi\n" );  /* remember address */
 831
 832     /* Set up temporary area */
 833
 834     if ( !isEx )
 835     {
 836         fprintf( outfile, "\tleal 4(%%edi), %%edi\n" );
 837
 838         fprintf( outfile, "\tleal -8(%%esp), %%eax\n" );
 839         fprintf( outfile, "\tmovl %%eax, -8(%%edi)\n" );    /* 32-bit sp */
 840
 841         fprintf( outfile, "\tmovw %%ss, %%ax\n" );
 842         fprintf( outfile, "\tandl $0x0000ffff, %%eax\n" );
 843         fprintf( outfile, "\tmovl %%eax, -4(%%edi)\n" );    /* 32-bit ss */
 844
 845         fprintf( outfile, "\taddl $%d, %%ebx\n", sizeof(STACK16FRAME)-size+4 + 4 );
 846         fprintf( outfile, "\tmovl %%ebx, 0(%%edi)\n" );    /* 16-bit ss:sp */
 847
 848         fprintf( outfile, "\tmovl " PREFIX "CALL32_%s_RetAddr, %%eax\n", name );
 849         fprintf( outfile, "\tmovl %%eax, 4(%%edi)\n" );   /* overwrite return address */
 850     }
 851     else
 852     {
 853         fprintf( outfile, "\taddl $%d, %%ebx\n", sizeof(STACK16FRAME)-size+4 );
 854         fprintf( outfile, "\tmovl %%ebx, 0(%%edi)\n" );
 855
 856         fprintf( outfile, "\tmovw %%ds, %%ax\n" );
 857         fprintf( outfile, "\tmovw %%ax, 4(%%edi)\n" );
 858
 859         fprintf( outfile, "\taddl $20, %%ebx\n" );
 860         fprintf( outfile, "\tmovw %%bx, 10(%%edi)\n" );
 861
 862         fprintf( outfile, "\tleal -8(%%esp), %%eax\n" );
 863         fprintf( outfile, "\tmovl %%eax, 12(%%edi)\n" );
 864
 865         fprintf( outfile, "\tmovw %%ss, %%ax\n" );
 866         fprintf( outfile, "\tandl $0x0000ffff, %%eax\n" );
 867         fprintf( outfile, "\tmovl %%eax, 16(%%edi)\n" );
 868
 869         fprintf( outfile, "\tmovl " PREFIX "CALL32_%s_RetAddr, %%eax\n", name );
 870         fprintf( outfile, "\tmovl %%eax, 20(%%edi)\n" );
 871     }
 872
 873     /* Set up registers and call CBClient relay stub (simulating a far call) */
 874
 875     fprintf( outfile, "\tmovl 16(%%ebp), %%esi\n" );
 876     fprintf( outfile, "\tmovl (%%esi), %%esi\n" );
 877
 878     fprintf( outfile, "\tmovl %%edi, %%ebx\n" );
 879     fprintf( outfile, "\tmovl 8(%%ebp), %%eax\n" );
 880     fprintf( outfile, "\tmovl 12(%%ebp), %%ebp\n" );
 881
 882     fprintf( outfile, "\tpushl %%cs\n" );
 883     fprintf( outfile, "\tcall *%%eax\n" );
 884
 885     /* Return new esi value to caller */
 886
 887     fprintf( outfile, "\tmovl 32(%%esp), %%edi\n" );
 888     fprintf( outfile, "\tmovl %%esi, (%%edi)\n" );
 889
 890     /* Cleanup temporary area (simulate STACK16_POP) */
 891
 892     fprintf( outfile, "\tpop %%esi\n" );
 893
 894     fprintf( outfile, "\tpushf\n" );
 895     fprintf( outfile, "\tstd\n" );
 896     fprintf( outfile, "\tdec %%esi\n" );
 897     fprintf( outfile, "\tleal %d(%%esi), %%edi\n", size );
 898     fprintf( outfile, "\tmovl $%d, %%ecx\n", sizeof(STACK16FRAME) );
 899     fprintf( outfile, "\trep\n\tmovsb\n" );
 900     fprintf( outfile, "\tpopf\n" );
 901
 902     fprintf( outfile, "\t.byte 0x64\n\taddw $%d,(%d)\n", size, STACKOFFSET );
 903
 904     /* Return argument size to caller */
 905     if ( isEx )
 906     {
 907         fprintf( outfile, "\tmovl 32(%%esp), %%ebx\n" );
 908         fprintf( outfile, "\tmovl %%ebp, (%%ebx)\n" );
 909     }
 910
 911     /* Restore registers and return */
 912
 913     fprintf( outfile, "\tpopl %%ebx\n" );
 914     fprintf( outfile, "\tpopl %%esi\n" );
 915     fprintf( outfile, "\tpopl %%edi\n" );
 916     fprintf( outfile, "\tpopl %%ebp\n" );
 917     fprintf( outfile, "\tret\n" );
 918 }
 919
 920 static void BuildCallTo32CBClientRet( FILE *outfile, BOOL isEx )
 921 {
 922     char *name = isEx? "CBClientEx" : "CBClient";
 923
 924     /* '16-bit' return stub */
 925
 926     fprintf( outfile, "\n\t.globl " PREFIX "CALL32_%s_Ret\n", name );
 927     fprintf( outfile, PREFIX "CALL32_%s_Ret:\n", name );
 928
 929     if ( !isEx )
 930     {
 931         fprintf( outfile, "\tmovzwl %%sp, %%ebx\n" );
 932         fprintf( outfile, "\tlssl %%ss:-16(%%ebx), %%esp\n" );
 933     }
 934     else
 935     {
 936         fprintf( outfile, "\tmovzwl %%bp, %%ebx\n" );
 937         fprintf( outfile, "\tsubw %%bp, %%sp\n" );
 938         fprintf( outfile, "\tmovzwl %%sp, %%ebp\n" );
 939         fprintf( outfile, "\tlssl %%ss:-12(%%ebx), %%esp\n" );
 940     }
 941     fprintf( outfile, "\tlret\n" );
 942
 943     /* Declare the return address variable */
 944
 945     fprintf( outfile, "\n\t.globl " PREFIX "CALL32_%s_RetAddr\n", name );
 946     fprintf( outfile, PREFIX "CALL32_%s_RetAddr:\t.long 0\n", name );
 947 }
 948
 949
 950 /*******************************************************************
 951  *         BuildCallFrom32Regs
 952  *
 953  * Build a 32-bit-to-Wine call-back function for a 'register' function.
 954  * 'args' is the number of dword arguments.
 955  *
 956  * Stack layout:
 957  *   ...
 958  * (ebp+12)  first arg
 959  * (ebp+8)   ret addr to user code
 960  * (ebp+4)   ret addr to relay code
 961  * (ebp+0)   saved ebp
 962  * (ebp-128) buffer area to allow stack frame manipulation
 963  * (ebp-332) CONTEXT86 struct
 964  * (ebp-336) CONTEXT86 *argument
 965  *  ....     other arguments copied from (ebp+12)
 966  *
 967  * The entry point routine is called with a CONTEXT* extra argument,
 968  * following the normal args. In this context structure, EIP_reg
 969  * contains the return address to user code, and ESP_reg the stack
 970  * pointer on return (with the return address and arguments already
 971  * removed).
 972  */
 973 static void BuildCallFrom32Regs( FILE *outfile )
 974 {
 975     static const int STACK_SPACE = 128 + sizeof(CONTEXT86);
 976
 977     /* Function header */
 978
 979     fprintf( outfile, "\n\t.align 4\n" );
 980 #ifdef USE_STABS
 981     fprintf( outfile, ".stabs \"CALL32_Regs:F1\",36,0,0," PREFIX "CALL32_Regs\n" );
 982 #endif
 983     fprintf( outfile, "\t.globl " PREFIX "CALL32_Regs\n" );
 984     fprintf( outfile, PREFIX "CALL32_Regs:\n" );
 985
 986     /* Allocate some buffer space on the stack */
 987
 988     fprintf( outfile, "\tpushl %%ebp\n" );
 989     fprintf( outfile, "\tmovl %%esp,%%ebp\n ");
 990     fprintf( outfile, "\tleal -%d(%%esp), %%esp\n", STACK_SPACE );
 991
 992     /* Build the context structure */
 993
 994     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Eax) - STACK_SPACE );
 995     fprintf( outfile, "\tpushfl\n" );
 996     fprintf( outfile, "\tpopl %%eax\n" );
 997     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(EFlags) - STACK_SPACE );
 998     fprintf( outfile, "\tmovl 0(%%ebp),%%eax\n" );
 999     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Ebp) - STACK_SPACE );
1000     fprintf( outfile, "\tmovl %%ebx,%d(%%ebp)\n", CONTEXTOFFSET(Ebx) - STACK_SPACE );
1001     fprintf( outfile, "\tmovl %%ecx,%d(%%ebp)\n", CONTEXTOFFSET(Ecx) - STACK_SPACE );
1002     fprintf( outfile, "\tmovl %%edx,%d(%%ebp)\n", CONTEXTOFFSET(Edx) - STACK_SPACE );
1003     fprintf( outfile, "\tmovl %%esi,%d(%%ebp)\n", CONTEXTOFFSET(Esi) - STACK_SPACE );
1004     fprintf( outfile, "\tmovl %%edi,%d(%%ebp)\n", CONTEXTOFFSET(Edi) - STACK_SPACE );
1005
1006     fprintf( outfile, "\txorl %%eax,%%eax\n" );
1007     fprintf( outfile, "\tmovw %%cs,%%ax\n" );
1008     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegCs) - STACK_SPACE );
1009     fprintf( outfile, "\tmovw %%es,%%ax\n" );
1010     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegEs) - STACK_SPACE );
1011     fprintf( outfile, "\tmovw %%fs,%%ax\n" );
1012     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegFs) - STACK_SPACE );
1013     fprintf( outfile, "\tmovw %%gs,%%ax\n" );
1014     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegGs) - STACK_SPACE );
1015     fprintf( outfile, "\tmovw %%ss,%%ax\n" );
1016     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegSs) - STACK_SPACE );
1017     fprintf( outfile, "\tmovw %%ds,%%ax\n" );
1018     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegDs) - STACK_SPACE );
1019     fprintf( outfile, "\tmovw %%ax,%%es\n" );  /* set %es equal to %ds just in case */
1020
1021     fprintf( outfile, "\tmovl $0x%x,%%eax\n", CONTEXT86_FULL );
1022     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(ContextFlags) - STACK_SPACE );
1023
1024     fprintf( outfile, "\tmovl 8(%%ebp),%%eax\n" ); /* Get %eip at time of call */
1025     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Eip) - STACK_SPACE );
1026
1027     /* Transfer the arguments */
1028
1029     fprintf( outfile, "\tmovl 4(%%ebp),%%ebx\n" );   /* get relay code addr */
1030     fprintf( outfile, "\tpushl %%esp\n" );           /* push ptr to context struct */
1031     fprintf( outfile, "\tmovzbl 4(%%ebx),%%ecx\n" ); /* fetch number of args to copy */
1032     fprintf( outfile, "\tjecxz 1f\n" );
1033     fprintf( outfile, "\tsubl %%ecx,%%esp\n" );
1034     fprintf( outfile, "\tleal 12(%%ebp),%%esi\n" );  /* get %esp at time of call */
1035     fprintf( outfile, "\tmovl %%esp,%%edi\n" );
1036     fprintf( outfile, "\tshrl $2,%%ecx\n" );
1037     fprintf( outfile, "\tcld\n" );
1038     fprintf( outfile, "\trep\n\tmovsl\n" );  /* copy args */
1039
1040     fprintf( outfile, "1:\tmovzbl 5(%%ebx),%%eax\n" ); /* fetch number of args to remove */
1041     fprintf( outfile, "\tleal 12(%%ebp,%%eax),%%eax\n" );
1042     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Esp) - STACK_SPACE );
1043
1044     /* Call the entry point */
1045
1046     fprintf( outfile, "\tcall *0(%%ebx)\n" );
1047
1048     /* Store %eip and %ebp onto the new stack */
1049
1050     fprintf( outfile, "\tmovl %d(%%ebp),%%edx\n", CONTEXTOFFSET(Esp) - STACK_SPACE );
1051     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(Eip) - STACK_SPACE );
1052     fprintf( outfile, "\tmovl %%eax,-4(%%edx)\n" );
1053     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(Ebp) - STACK_SPACE );
1054     fprintf( outfile, "\tmovl %%eax,-8(%%edx)\n" );
1055
1056     /* Restore the context structure */
1057
1058     /* Note: we don't bother to restore %cs, %ds and %ss
1059      *       changing them in 32-bit code is a recipe for disaster anyway
1060      */
1061     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(SegEs) - STACK_SPACE );
1062     fprintf( outfile, "\tmovw %%ax,%%es\n" );
1063     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(SegFs) - STACK_SPACE );
1064     fprintf( outfile, "\tmovw %%ax,%%fs\n" );
1065     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(SegGs) - STACK_SPACE );
1066     fprintf( outfile, "\tmovw %%ax,%%gs\n" );
1067
1068     fprintf( outfile, "\tmovl %d(%%ebp),%%edi\n", CONTEXTOFFSET(Edi) - STACK_SPACE );
1069     fprintf( outfile, "\tmovl %d(%%ebp),%%esi\n", CONTEXTOFFSET(Esi) - STACK_SPACE );
1070     fprintf( outfile, "\tmovl %d(%%ebp),%%edx\n", CONTEXTOFFSET(Edx) - STACK_SPACE );
1071     fprintf( outfile, "\tmovl %d(%%ebp),%%ecx\n", CONTEXTOFFSET(Ecx) - STACK_SPACE );
1072     fprintf( outfile, "\tmovl %d(%%ebp),%%ebx\n", CONTEXTOFFSET(Ebx) - STACK_SPACE );
1073
1074     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(EFlags) - STACK_SPACE );
1075     fprintf( outfile, "\tpushl %%eax\n" );
1076     fprintf( outfile, "\tpopfl\n" );
1077     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(Eax) - STACK_SPACE );
1078
1079     fprintf( outfile, "\tmovl %d(%%ebp),%%ebp\n", CONTEXTOFFSET(Esp) - STACK_SPACE );
1080     fprintf( outfile, "\tleal -8(%%ebp),%%esp\n" );
1081     fprintf( outfile, "\tpopl %%ebp\n" );
1082     fprintf( outfile, "\tret\n" );
1083 }
1084
1085
1086 /*******************************************************************
1087  *         BuildRelays
1088  *
1089  * Build all the relay callbacks
1090  */
1091 void BuildRelays( FILE *outfile )
1092 {
1093     /* File header */
1094
1095     fprintf( outfile, "/* File generated automatically. Do not edit! */\n\n" );
1096     fprintf( outfile, "\t.text\n" );
1097
1098 #ifdef __i386__
1099
1100 #ifdef USE_STABS
1101     if (output_file_name)
1102     {
1103         char buffer[1024];
1104         getcwd(buffer, sizeof(buffer));
1105         fprintf( outfile, "\t.file\t\"%s\"\n", output_file_name );
1106
1107         /*
1108          * The stabs help the internal debugger as they are an indication that it
1109          * is sensible to step into a thunk/trampoline.
1110          */
1111         fprintf( outfile, ".stabs \"%s/\",100,0,0,Code_Start\n", buffer);
1112         fprintf( outfile, ".stabs \"%s\",100,0,0,Code_Start\n", output_file_name );
1113         fprintf( outfile, "Code_Start:\n\n" );
1114     }
1115 #endif
1116     fprintf( outfile, PREFIX"Call16_Start:\n" );
1117     fprintf( outfile, "\t.globl "PREFIX"Call16_Start\n" );
1118     fprintf( outfile, "\t.byte 0\n\n" );
1119
1120     /* Standard CallFrom16 routine (WORD return) */
1121     BuildCallFrom16Core( outfile, FALSE, FALSE, TRUE );
1122
1123     /* Standard CallFrom16 routine (DWORD return) */
1124     BuildCallFrom16Core( outfile, FALSE, FALSE, FALSE );
1125
1126     /* Register CallFrom16 routine */
1127     BuildCallFrom16Core( outfile, TRUE, FALSE, FALSE );
1128
1129     /* C16ThkSL CallFrom16 routine */
1130     BuildCallFrom16Core( outfile, FALSE, TRUE, FALSE );
1131
1132     /* Standard CallTo16 routine (WORD return) */
1133     BuildCallTo16Core( outfile, TRUE, FALSE );
1134
1135     /* Standard CallTo16 routine (DWORD return) */
1136     BuildCallTo16Core( outfile, FALSE, FALSE );
1137
1138     /* Register CallTo16 routine (16:16 retf) */
1139     BuildCallTo16Core( outfile, FALSE, 1 );
1140
1141     /* Register CallTo16 routine (16:32 retf) */
1142     BuildCallTo16Core( outfile, FALSE, 2 );
1143
1144     /* CBClientThunkSL routine */
1145     BuildCallTo32CBClient( outfile, FALSE );
1146
1147     /* CBClientThunkSLEx routine */
1148     BuildCallTo32CBClient( outfile, TRUE  );
1149
1150     /* 32-bit register entry point */
1151     BuildCallFrom32Regs( outfile );
1152
1153     fprintf( outfile, PREFIX"Call16_End:\n" );
1154     fprintf( outfile, "\t.globl "PREFIX"Call16_End\n" );
1155
1156 #ifdef USE_STABS
1157     fprintf( outfile, "\t.stabs \"\",100,0,0,.Letext\n");
1158     fprintf( outfile, ".Letext:\n");
1159 #endif
1160
1161     /* The whole Call16_Ret segment must lie within the .data section */
1162     fprintf( outfile, "\n\t.data\n" );
1163     fprintf( outfile, "\t.globl " PREFIX "Call16_Ret_Start\n" );
1164     fprintf( outfile, PREFIX "Call16_Ret_Start:\n" );
1165
1166     /* Standard CallTo16 return stub */
1167     BuildRet16Func( outfile );
1168
1169     /* CBClientThunkSL return stub */
1170     BuildCallTo32CBClientRet( outfile, FALSE );
1171
1172     /* CBClientThunkSLEx return stub */
1173     BuildCallTo32CBClientRet( outfile, TRUE  );
1174
1175     /* End of Call16_Ret segment */
1176     fprintf( outfile, "\n\t.globl " PREFIX "Call16_Ret_End\n" );
1177     fprintf( outfile, PREFIX "Call16_Ret_End:\n" );
1178
1179 #else  /* __i386__ */
1180
1181     fprintf( outfile, PREFIX"Call16_Start:\n" );
1182     fprintf( outfile, "\t.globl "PREFIX"Call16_Start\n" );
1183     fprintf( outfile, "\t.byte 0\n\n" );
1184     fprintf( outfile, PREFIX"Call16_End:\n" );
1185     fprintf( outfile, "\t.globl "PREFIX"Call16_End\n" );
1186
1187     fprintf( outfile, "\t.globl " PREFIX "Call16_Ret_Start\n" );
1188     fprintf( outfile, PREFIX "Call16_Ret_Start:\n" );
1189     fprintf( outfile, "\t.byte 0\n\n" );
1190     fprintf( outfile, "\n\t.globl " PREFIX "Call16_Ret_End\n" );
1191     fprintf( outfile, PREFIX "Call16_Ret_End:\n" );
1192
1193 #endif  /* __i386__ */
1194 }