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? "regs" : short_ret? "word" : "long";
  91
  92     /* Function header */
  93     fprintf( outfile, "\n\t.align 4\n" );
  94 #ifdef USE_STABS
  95     fprintf( outfile, ".stabs \"__wine_call_from_16_%s:F1\",36,0,0," PREFIX "__wine_call_from_16_%s\n", name, name);
  96 #endif
  97     fprintf( outfile, "\t.type " PREFIX "__wine_call_from_16_%s,@function\n", name );
  98     fprintf( outfile, "\t.globl " PREFIX "__wine_call_from_16_%s\n", name );
  99     fprintf( outfile, PREFIX "__wine_call_from_16_%s:\n", name );
 100
 101     /* Create STACK16FRAME (except STACK32FRAME link) */
 102     fprintf( outfile, "\tpushw %%gs\n" );
 103     fprintf( outfile, "\tpushw %%fs\n" );
 104     fprintf( outfile, "\tpushw %%es\n" );
 105     fprintf( outfile, "\tpushw %%ds\n" );
 106     fprintf( outfile, "\tpushl %%ebp\n" );
 107     fprintf( outfile, "\tpushl %%ecx\n" );
 108     fprintf( outfile, "\tpushl %%edx\n" );
 109
 110     /* Save original EFlags register */
 111     fprintf( outfile, "\tpushfl\n" );
 112
 113     if ( UsePIC )
 114     {
 115         /* Get Global Offset Table into %ecx */
 116         fprintf( outfile, "\tcall .L__wine_call_from_16_%s.getgot1\n", name );
 117         fprintf( outfile, ".L__wine_call_from_16_%s.getgot1:\n", name );
 118         fprintf( outfile, "\tpopl %%ecx\n" );
 119         fprintf( outfile, "\taddl $_GLOBAL_OFFSET_TABLE_+[.-.L__wine_call_from_16_%s.getgot1], %%ecx\n", name );
 120     }
 121
 122     if (UsePIC)
 123     {
 124         fprintf( outfile, "\t.byte 0x2e\n\tmovl " PREFIX "CallTo16_DataSelector@GOT(%%ecx), %%edx\n" );
 125         fprintf( outfile, "\t.byte 0x2e\n\tmovl (%%edx), %%edx\n" );
 126     }
 127     else
 128         fprintf( outfile, "\t.byte 0x2e\n\tmovl " PREFIX "CallTo16_DataSelector,%%edx\n" );
 129
 130     /* Load 32-bit segment registers */
 131 #ifdef __svr4__
 132     fprintf( outfile, "\tdata16\n");
 133 #endif
 134     fprintf( outfile, "\tmovw %%dx, %%ds\n" );
 135 #ifdef __svr4__
 136     fprintf( outfile, "\tdata16\n");
 137 #endif
 138     fprintf( outfile, "\tmovw %%dx, %%es\n" );
 139
 140     if ( UsePIC )
 141     {
 142         fprintf( outfile, "\tmovl " PREFIX "SYSLEVEL_Win16CurrentTeb@GOT(%%ecx), %%edx\n" );
 143         fprintf( outfile, "\tmovw (%%edx), %%fs\n" );
 144     }
 145     else
 146         fprintf( outfile, "\tmovw " PREFIX "SYSLEVEL_Win16CurrentTeb, %%fs\n" );
 147
 148     /* Get address of ldt_copy array into %ecx */
 149     if ( UsePIC )
 150         fprintf( outfile, "\tmovl " PREFIX "ldt_copy@GOT(%%ecx), %%ecx\n" );
 151     else
 152         fprintf( outfile, "\tmovl $" PREFIX "ldt_copy, %%ecx\n" );
 153
 154     /* Translate STACK16FRAME base to flat offset in %edx */
 155     fprintf( outfile, "\tmovw %%ss, %%dx\n" );
 156     fprintf( outfile, "\tandl $0xfff8, %%edx\n" );
 157     fprintf( outfile, "\tmovl (%%ecx,%%edx), %%edx\n" );
 158     fprintf( outfile, "\tmovzwl %%sp, %%ebp\n" );
 159     fprintf( outfile, "\tleal (%%ebp,%%edx), %%edx\n" );
 160
 161     /* Get saved flags into %ecx */
 162     fprintf( outfile, "\tpopl %%ecx\n" );
 163
 164     /* Get the 32-bit stack pointer from the TEB and complete STACK16FRAME */
 165     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d), %%ebp\n", STACKOFFSET );
 166     fprintf( outfile, "\tpushl %%ebp\n" );
 167
 168     /* Switch stacks */
 169 #ifdef __svr4__
 170     fprintf( outfile,"\tdata16\n");
 171 #endif
 172     fprintf( outfile, "\t.byte 0x64\n\tmovw %%ss, (%d)\n", STACKOFFSET + 2 );
 173     fprintf( outfile, "\t.byte 0x64\n\tmovw %%sp, (%d)\n", STACKOFFSET );
 174     fprintf( outfile, "\tpushl %%ds\n" );
 175     fprintf( outfile, "\tpopl %%ss\n" );
 176     fprintf( outfile, "\tmovl %%ebp, %%esp\n" );
 177     fprintf( outfile, "\taddl $%d, %%ebp\n", STRUCTOFFSET(STACK32FRAME, ebp) );
 178
 179
 180     /* At this point:
 181        STACK16FRAME is completely set up
 182        DS, ES, SS: flat data segment
 183        FS: current TEB
 184        ESP: points to last STACK32FRAME
 185        EBP: points to ebp member of last STACK32FRAME
 186        EDX: points to current STACK16FRAME
 187        ECX: contains saved flags
 188        all other registers: unchanged */
 189
 190     /* Special case: C16ThkSL stub */
 191     if ( thunk )
 192     {
 193         /* Set up registers as expected and call thunk */
 194         fprintf( outfile, "\tleal %d(%%edx), %%ebx\n", sizeof(STACK16FRAME)-22 );
 195         fprintf( outfile, "\tleal -4(%%esp), %%ebp\n" );
 196
 197         fprintf( outfile, "\tcall *%d(%%edx)\n", STACK16OFFSET(entry_point) );
 198
 199         /* Switch stack back */
 200         /* fprintf( outfile, "\t.byte 0x64\n\tlssw (%d), %%sp\n", STACKOFFSET ); */
 201         fprintf( outfile, "\t.byte 0x64,0x66,0x0f,0xb2,0x25\n\t.long %d\n", STACKOFFSET );
 202         fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 203
 204         /* Restore registers and return directly to caller */
 205         fprintf( outfile, "\taddl $8, %%esp\n" );
 206         fprintf( outfile, "\tpopl %%ebp\n" );
 207         fprintf( outfile, "\tpopw %%ds\n" );
 208         fprintf( outfile, "\tpopw %%es\n" );
 209         fprintf( outfile, "\tpopw %%fs\n" );
 210         fprintf( outfile, "\tpopw %%gs\n" );
 211         fprintf( outfile, "\taddl $20, %%esp\n" );
 212
 213         fprintf( outfile, "\txorb %%ch, %%ch\n" );
 214         fprintf( outfile, "\tpopl %%ebx\n" );
 215         fprintf( outfile, "\taddw %%cx, %%sp\n" );
 216         fprintf( outfile, "\tpush %%ebx\n" );
 217
 218         fprintf( outfile, "\t.byte 0x66\n" );
 219         fprintf( outfile, "\tlret\n" );
 220
 221         return;
 222     }
 223
 224
 225     /* Build register CONTEXT */
 226     if ( reg_func )
 227     {
 228         fprintf( outfile, "\tsubl $%d, %%esp\n", sizeof(CONTEXT86) );
 229
 230         fprintf( outfile, "\tmovl %%ecx, %d(%%esp)\n", CONTEXTOFFSET(EFlags) );
 231
 232         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Eax) );
 233         fprintf( outfile, "\tmovl %%ebx, %d(%%esp)\n", CONTEXTOFFSET(Ebx) );
 234         fprintf( outfile, "\tmovl %%esi, %d(%%esp)\n", CONTEXTOFFSET(Esi) );
 235         fprintf( outfile, "\tmovl %%edi, %d(%%esp)\n", CONTEXTOFFSET(Edi) );
 236
 237         fprintf( outfile, "\tmovl %d(%%edx), %%eax\n", STACK16OFFSET(ebp) );
 238         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Ebp) );
 239         fprintf( outfile, "\tmovl %d(%%edx), %%eax\n", STACK16OFFSET(ecx) );
 240         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Ecx) );
 241         fprintf( outfile, "\tmovl %d(%%edx), %%eax\n", STACK16OFFSET(edx) );
 242         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Edx) );
 243
 244         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(ds) );
 245         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegDs) );
 246         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(es) );
 247         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegEs) );
 248         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(fs) );
 249         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegFs) );
 250         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(gs) );
 251         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegGs) );
 252
 253         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(cs) );
 254         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegCs) );
 255         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(ip) );
 256         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Eip) );
 257
 258         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%eax\n", STACKOFFSET+2 );
 259         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegSs) );
 260         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%eax\n", STACKOFFSET );
 261         fprintf( outfile, "\taddl $%d, %%eax\n", STACK16OFFSET(ip) );
 262         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Esp) );
 263 #if 0
 264         fprintf( outfile, "\tfsave %d(%%esp)\n", CONTEXTOFFSET(FloatSave) );
 265 #endif
 266
 267         /* Push address of CONTEXT86 structure -- popped by the relay routine */
 268         fprintf( outfile, "\tpushl %%esp\n" );
 269     }
 270
 271
 272     /* Print debug info before call */
 273     if ( debugging )
 274     {
 275         if ( UsePIC )
 276         {
 277             fprintf( outfile, "\tpushl %%ebx\n" );
 278
 279             /* Get Global Offset Table into %ebx (for PLT call) */
 280             fprintf( outfile, "\tcall .L__wine_call_from_16_%s.getgot2\n", name );
 281             fprintf( outfile, ".L__wine_call_from_16_%s.getgot2:\n", name );
 282             fprintf( outfile, "\tpopl %%ebx\n" );
 283             fprintf( outfile, "\taddl $_GLOBAL_OFFSET_TABLE_+[.-.L__wine_call_from_16_%s.getgot2], %%ebx\n", name );
 284         }
 285
 286         fprintf( outfile, "\tpushl %%edx\n" );
 287         if ( reg_func )
 288             fprintf( outfile, "\tleal -%d(%%ebp), %%eax\n\tpushl %%eax\n",
 289                               sizeof(CONTEXT) + STRUCTOFFSET(STACK32FRAME, ebp) );
 290         else
 291             fprintf( outfile, "\tpushl $0\n" );
 292
 293         if ( UsePIC )
 294             fprintf( outfile, "\tcall " PREFIX "RELAY_DebugCallFrom16@PLT\n ");
 295         else
 296             fprintf( outfile, "\tcall " PREFIX "RELAY_DebugCallFrom16\n ");
 297
 298         fprintf( outfile, "\tpopl %%edx\n" );
 299         fprintf( outfile, "\tpopl %%edx\n" );
 300
 301         if ( UsePIC )
 302             fprintf( outfile, "\tpopl %%ebx\n" );
 303     }
 304
 305     /* Call relay routine (which will call the API entry point) */
 306     fprintf( outfile, "\tleal %d(%%edx), %%eax\n", sizeof(STACK16FRAME) );
 307     fprintf( outfile, "\tpushl %%eax\n" );
 308     fprintf( outfile, "\tpushl %d(%%edx)\n", STACK16OFFSET(entry_point) );
 309     fprintf( outfile, "\tcall *%d(%%edx)\n", STACK16OFFSET(relay) );
 310
 311     /* Print debug info after call */
 312     if ( debugging )
 313     {
 314         if ( UsePIC )
 315         {
 316             fprintf( outfile, "\tpushl %%ebx\n" );
 317
 318             /* Get Global Offset Table into %ebx (for PLT call) */
 319             fprintf( outfile, "\tcall .L__wine_call_from_16_%s.getgot3\n", name );
 320             fprintf( outfile, ".L__wine_call_from_16_%s.getgot3:\n", name );
 321             fprintf( outfile, "\tpopl %%ebx\n" );
 322             fprintf( outfile, "\taddl $_GLOBAL_OFFSET_TABLE_+[.-.L__wine_call_from_16_%s.getgot3], %%ebx\n", name );
 323         }
 324
 325         fprintf( outfile, "\tpushl %%eax\n" );
 326         if ( reg_func )
 327             fprintf( outfile, "\tleal -%d(%%ebp), %%eax\n\tpushl %%eax\n",
 328                               sizeof(CONTEXT) + STRUCTOFFSET(STACK32FRAME, ebp) );
 329         else
 330             fprintf( outfile, "\tpushl $0\n" );
 331
 332         if ( UsePIC )
 333             fprintf( outfile, "\tcall " PREFIX "RELAY_DebugCallFrom16Ret@PLT\n ");
 334         else
 335             fprintf( outfile, "\tcall " PREFIX "RELAY_DebugCallFrom16Ret\n ");
 336
 337         fprintf( outfile, "\tpopl %%eax\n" );
 338         fprintf( outfile, "\tpopl %%eax\n" );
 339
 340         if ( UsePIC )
 341             fprintf( outfile, "\tpopl %%ebx\n" );
 342     }
 343
 344
 345     if ( reg_func )
 346     {
 347         fprintf( outfile, "\tmovl %%esp, %%ebx\n" );
 348
 349         /* Switch stack back */
 350         /* fprintf( outfile, "\t.byte 0x64\n\tlssw (%d), %%sp\n", STACKOFFSET ); */
 351         fprintf( outfile, "\t.byte 0x64,0x66,0x0f,0xb2,0x25\n\t.long %d\n", STACKOFFSET );
 352         fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 353
 354         /* Get return address to CallFrom16 stub */
 355         fprintf( outfile, "\taddw $%d, %%sp\n", STACK16OFFSET(callfrom_ip)-4 );
 356         fprintf( outfile, "\tpopl %%eax\n" );
 357         fprintf( outfile, "\tpopl %%edx\n" );
 358
 359         /* Restore all registers from CONTEXT */
 360         fprintf( outfile, "\tmovw %d(%%ebx), %%ss\n", CONTEXTOFFSET(SegSs) );
 361         fprintf( outfile, "\tmovl %d(%%ebx), %%esp\n", CONTEXTOFFSET(Esp) );
 362         fprintf( outfile, "\taddl $4, %%esp\n" );  /* room for final return address */
 363
 364         fprintf( outfile, "\tpushw %d(%%ebx)\n", CONTEXTOFFSET(SegCs) );
 365         fprintf( outfile, "\tpushw %d(%%ebx)\n", CONTEXTOFFSET(Eip) );
 366         fprintf( outfile, "\tpushl %%edx\n" );
 367         fprintf( outfile, "\tpushl %%eax\n" );
 368         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(EFlags) );
 369         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(SegDs) );
 370
 371         fprintf( outfile, "\tmovw %d(%%ebx), %%es\n", CONTEXTOFFSET(SegEs) );
 372         fprintf( outfile, "\tmovw %d(%%ebx), %%fs\n", CONTEXTOFFSET(SegFs) );
 373         fprintf( outfile, "\tmovw %d(%%ebx), %%gs\n", CONTEXTOFFSET(SegGs) );
 374
 375         fprintf( outfile, "\tmovl %d(%%ebx), %%ebp\n", CONTEXTOFFSET(Ebp) );
 376         fprintf( outfile, "\tmovl %d(%%ebx), %%esi\n", CONTEXTOFFSET(Esi) );
 377         fprintf( outfile, "\tmovl %d(%%ebx), %%edi\n", CONTEXTOFFSET(Edi) );
 378         fprintf( outfile, "\tmovl %d(%%ebx), %%eax\n", CONTEXTOFFSET(Eax) );
 379         fprintf( outfile, "\tmovl %d(%%ebx), %%edx\n", CONTEXTOFFSET(Edx) );
 380         fprintf( outfile, "\tmovl %d(%%ebx), %%ecx\n", CONTEXTOFFSET(Ecx) );
 381         fprintf( outfile, "\tmovl %d(%%ebx), %%ebx\n", CONTEXTOFFSET(Ebx) );
 382
 383         fprintf( outfile, "\tpopl %%ds\n" );
 384         fprintf( outfile, "\tpopfl\n" );
 385         fprintf( outfile, "\tlret\n" );
 386     }
 387     else
 388     {
 389         /* Switch stack back */
 390         /* fprintf( outfile, "\t.byte 0x64\n\tlssw (%d), %%sp\n", STACKOFFSET ); */
 391         fprintf( outfile, "\t.byte 0x64,0x66,0x0f,0xb2,0x25\n\t.long %d\n", STACKOFFSET );
 392         fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 393
 394         /* Restore registers */
 395         fprintf( outfile, "\tpopl %%edx\n" );
 396         fprintf( outfile, "\tpopl %%ecx\n" );
 397         fprintf( outfile, "\tpopl %%ebp\n" );
 398         fprintf( outfile, "\tpopw %%ds\n" );
 399         fprintf( outfile, "\tpopw %%es\n" );
 400         fprintf( outfile, "\tpopw %%fs\n" );
 401         fprintf( outfile, "\tpopw %%gs\n" );
 402
 403         /* Prepare return value and set flags accordingly */
 404         if ( !short_ret )
 405             fprintf( outfile, "\tshldl $16, %%eax, %%edx\n" );
 406         fprintf( outfile, "\torl %%eax, %%eax\n" );
 407
 408         /* Return to return stub which will return to caller */
 409         fprintf( outfile, "\tlret $12\n" );
 410     }
 411 }
 412
 413
 414 /*******************************************************************
 415  *         BuildCallTo16Core
 416  *
 417  * This routine builds the core routines used in 32->16 thunks:
 418  *
 419  *   extern void CALLBACK CallTo16Word( SEGPTR target, int nb_args );
 420  *   extern void CALLBACK CallTo16Long( SEGPTR target, int nb_args );
 421  *   extern void CALLBACK CallTo16RegisterShort( const CONTEXT86 *context, int nb_args );
 422  *   extern void CALLBACK CallTo16RegisterLong ( const CONTEXT86 *context, int nb_args );
 423  *
 424  * These routines can be called directly from 32-bit code.
 425  *
 426  * All routines expect that the 16-bit stack contents (arguments) were
 427  * already set up by the caller; nb_args must contain the number of bytes
 428  * to be conserved.  The 16-bit SS:SP will be set accordinly.
 429  *
 430  * All other registers are either taken from the CONTEXT86 structure
 431  * or else set to default values.  The target routine address is either
 432  * given directly or taken from the CONTEXT86.
 433  *
 434  * If you want to call a 16-bit routine taking only standard argument types
 435  * (WORD and LONG), you can also have an appropriate argument conversion
 436  * stub automatically generated (see BuildCallTo16); you'd then call this
 437  * stub, which in turn would prepare the 16-bit stack and call the appropiate
 438  * core routine.
 439  *
 440  */
 441
 442 static void BuildCallTo16Core( FILE *outfile, int short_ret, int reg_func )
 443 {
 444     char *name = reg_func == 2 ? "RegisterLong" :
 445                  reg_func == 1 ? "RegisterShort" :
 446                  short_ret? "Word" : "Long";
 447
 448     /* Function header */
 449     fprintf( outfile, "\n\t.align 4\n" );
 450 #ifdef USE_STABS
 451     fprintf( outfile, ".stabs \"CallTo16%s:F1\",36,0,0," PREFIX "CallTo16%s\n",
 452              name, name);
 453 #endif
 454     fprintf( outfile, "\t.type " PREFIX "CallTo16%s,@function\n", name );
 455     fprintf( outfile, "\t.globl " PREFIX "CallTo16%s\n", name );
 456     fprintf( outfile, PREFIX "CallTo16%s:\n", name );
 457
 458     /* Function entry sequence */
 459     fprintf( outfile, "\tpushl %%ebp\n" );
 460     fprintf( outfile, "\tmovl %%esp, %%ebp\n" );
 461
 462     /* Save the 32-bit registers */
 463     fprintf( outfile, "\tpushl %%ebx\n" );
 464     fprintf( outfile, "\tpushl %%ecx\n" );
 465     fprintf( outfile, "\tpushl %%edx\n" );
 466     fprintf( outfile, "\tpushl %%esi\n" );
 467     fprintf( outfile, "\tpushl %%edi\n" );
 468
 469     if ( UsePIC )
 470     {
 471         /* Get Global Offset Table into %ebx */
 472         fprintf( outfile, "\tcall .LCallTo16%s.getgot1\n", name );
 473         fprintf( outfile, ".LCallTo16%s.getgot1:\n", name );
 474         fprintf( outfile, "\tpopl %%ebx\n" );
 475         fprintf( outfile, "\taddl $_GLOBAL_OFFSET_TABLE_+[.-.LCallTo16%s.getgot1], %%ebx\n", name );
 476     }
 477
 478     /* Enter Win16 Mutex */
 479     if ( UsePIC )
 480         fprintf( outfile, "\tcall " PREFIX "SYSLEVEL_EnterWin16Lock@PLT\n" );
 481     else
 482         fprintf( outfile, "\tcall " PREFIX "SYSLEVEL_EnterWin16Lock\n" );
 483
 484     /* Print debugging info */
 485     if (debugging)
 486     {
 487         /* Push flags, number of arguments, and target */
 488         fprintf( outfile, "\tpushl $%d\n", reg_func );
 489         fprintf( outfile, "\tpushl 12(%%ebp)\n" );
 490         fprintf( outfile, "\tpushl  8(%%ebp)\n" );
 491
 492         if ( UsePIC )
 493             fprintf( outfile, "\tcall " PREFIX "RELAY_DebugCallTo16@PLT\n" );
 494         else
 495             fprintf( outfile, "\tcall " PREFIX "RELAY_DebugCallTo16\n" );
 496
 497         fprintf( outfile, "\taddl $12, %%esp\n" );
 498     }
 499
 500     /* Get return address */
 501     if ( UsePIC )
 502     {
 503         fprintf( outfile, "\tmovl " PREFIX "CallTo16_RetAddr@GOT(%%ebx), %%ecx\n" );
 504         fprintf( outfile, "\tmovl " PREFIX "(%%ecx), %%ecx\n" );
 505     }
 506     else
 507         fprintf( outfile, "\tmovl " PREFIX "CallTo16_RetAddr, %%ecx\n" );
 508
 509     /* Call the actual CallTo16 routine (simulate a lcall) */
 510     fprintf( outfile, "\tpushl %%cs\n" );
 511     fprintf( outfile, "\tcall .LCallTo16%s\n", name );
 512
 513     if ( !reg_func )
 514     {
 515         /* Convert and push return value */
 516         if ( short_ret )
 517         {
 518             fprintf( outfile, "\tmovzwl %%ax, %%eax\n" );
 519             fprintf( outfile, "\tpushl %%eax\n" );
 520         }
 521         else
 522         {
 523             fprintf( outfile, "\tshll $16,%%edx\n" );
 524             fprintf( outfile, "\tmovw %%ax,%%dx\n" );
 525             fprintf( outfile, "\tpushl %%edx\n" );
 526         }
 527     }
 528     else
 529     {
 530         /*
 531          * Modify CONTEXT86 structure to contain new values
 532          *
 533          * NOTE:  We restore only EAX, EBX, EDX, EDX, EBP, and ESP.
 534          *        The segment registers as well as ESI and EDI should
 535          *        not be modified by a well-behaved 16-bit routine in
 536          *        any case.  [If necessary, we could restore them as well,
 537          *        at the cost of a somewhat less efficient return path.]
 538          */
 539
 540         fprintf( outfile, "\tmovl %d(%%esp), %%edi\n", STACK32OFFSET(target)-12 );
 541         fprintf( outfile, "\tmovl %%eax, %d(%%edi)\n", CONTEXTOFFSET(Eax) );
 542         fprintf( outfile, "\tmovl %%ebx, %d(%%edi)\n", CONTEXTOFFSET(Ebx) );
 543         fprintf( outfile, "\tmovl %%ecx, %d(%%edi)\n", CONTEXTOFFSET(Ecx) );
 544         fprintf( outfile, "\tmovl %%edx, %d(%%edi)\n", CONTEXTOFFSET(Edx) );
 545         fprintf( outfile, "\tmovl %%ebp, %d(%%edi)\n", CONTEXTOFFSET(Ebp) );
 546         fprintf( outfile, "\tmovl %%esi, %d(%%edi)\n", CONTEXTOFFSET(Esp) );
 547                  /* The return glue code saved %esp into %esi */
 548
 549         fprintf( outfile, "\tpushl %%edi\n" );
 550     }
 551
 552     if ( UsePIC )
 553     {
 554         /* Get Global Offset Table into %ebx (might have been overwritten) */
 555         fprintf( outfile, "\tcall .LCallTo16%s.getgot2\n", name );
 556         fprintf( outfile, ".LCallTo16%s.getgot2:\n", name );
 557         fprintf( outfile, "\tpopl %%ebx\n" );
 558         fprintf( outfile, "\taddl $_GLOBAL_OFFSET_TABLE_+[.-.LCallTo16%s.getgot2], %%ebx\n", name );
 559     }
 560
 561     /* Print debugging info */
 562     if (debugging)
 563     {
 564         fprintf( outfile, "\tpushl $%d\n", reg_func );
 565
 566         if ( UsePIC )
 567             fprintf( outfile, "\tcall " PREFIX "RELAY_DebugCallTo16Ret@PLT\n" );
 568         else
 569             fprintf( outfile, "\tcall " PREFIX "RELAY_DebugCallTo16Ret\n" );
 570
 571         fprintf( outfile, "\taddl $4, %%esp\n" );
 572     }
 573
 574     /* Leave Win16 Mutex */
 575     if ( UsePIC )
 576         fprintf( outfile, "\tcall " PREFIX "SYSLEVEL_LeaveWin16Lock@PLT\n" );
 577     else
 578         fprintf( outfile, "\tcall " PREFIX "SYSLEVEL_LeaveWin16Lock\n" );
 579
 580     /* Get return value */
 581     fprintf( outfile, "\tpopl %%eax\n" );
 582
 583     /* Restore the 32-bit registers */
 584     fprintf( outfile, "\tpopl %%edi\n" );
 585     fprintf( outfile, "\tpopl %%esi\n" );
 586     fprintf( outfile, "\tpopl %%edx\n" );
 587     fprintf( outfile, "\tpopl %%ecx\n" );
 588     fprintf( outfile, "\tpopl %%ebx\n" );
 589
 590     /* Function exit sequence */
 591     fprintf( outfile, "\tpopl %%ebp\n" );
 592     fprintf( outfile, "\tret $8\n" );
 593
 594
 595     /* Start of the actual CallTo16 routine */
 596
 597     fprintf( outfile, ".LCallTo16%s:\n", name );
 598
 599     /* Complete STACK32FRAME */
 600     fprintf( outfile, "\t.byte 0x64\n\tpushl (%d)\n", STACKOFFSET );
 601     fprintf( outfile, "\tmovl %%esp,%%edx\n" );
 602
 603     /* Switch to the 16-bit stack */
 604 #ifdef __svr4__
 605     fprintf( outfile,"\tdata16\n");
 606 #endif
 607     fprintf( outfile, "\t.byte 0x64\n\tmovw (%d),%%ss\n", STACKOFFSET + 2);
 608     fprintf( outfile, "\t.byte 0x64\n\tmovw (%d),%%sp\n", STACKOFFSET );
 609     fprintf( outfile, "\t.byte 0x64\n\tmovl %%edx,(%d)\n", STACKOFFSET );
 610
 611     /* Make %bp point to the previous stackframe (built by CallFrom16) */
 612     fprintf( outfile, "\tmovzwl %%sp,%%ebp\n" );
 613     fprintf( outfile, "\tleal %d(%%ebp),%%ebp\n", STACK16OFFSET(bp) );
 614
 615     /* Add the specified offset to the new sp */
 616     fprintf( outfile, "\tsubw %d(%%edx), %%sp\n", STACK32OFFSET(nb_args) );
 617
 618     /* Push the return address
 619      * With sreg suffix, we push 16:16 address (normal lret)
 620      * With lreg suffix, we push 16:32 address (0x66 lret, for KERNEL32_45)
 621      */
 622     if (reg_func != 2)
 623         fprintf( outfile, "\tpushl %%ecx\n" );
 624     else
 625     {
 626         fprintf( outfile, "\tshldl $16, %%ecx, %%eax\n" );
 627         fprintf( outfile, "\tpushw $0\n" );
 628         fprintf( outfile, "\tpushw %%ax\n" );
 629         fprintf( outfile, "\tpushw $0\n" );
 630         fprintf( outfile, "\tpushw %%cx\n" );
 631     }
 632
 633     if (reg_func)
 634     {
 635         /* Push the called routine address */
 636         fprintf( outfile, "\tmovl %d(%%edx),%%edx\n", STACK32OFFSET(target) );
 637         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(SegCs) );
 638         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(Eip) );
 639
 640         /* Get the registers */
 641         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(SegDs) );
 642         fprintf( outfile, "\tmovl %d(%%edx),%%eax\n", CONTEXTOFFSET(SegEs) );
 643         fprintf( outfile, "\tmovw %%ax,%%es\n" );
 644         fprintf( outfile, "\tmovl %d(%%edx),%%eax\n", CONTEXTOFFSET(SegFs) );
 645         fprintf( outfile, "\tmovw %%ax,%%fs\n" );
 646         fprintf( outfile, "\tmovl %d(%%edx),%%ebp\n", CONTEXTOFFSET(Ebp) );
 647         fprintf( outfile, "\tmovl %d(%%edx),%%esi\n", CONTEXTOFFSET(Esi) );
 648         fprintf( outfile, "\tmovl %d(%%edx),%%edi\n", CONTEXTOFFSET(Edi) );
 649         fprintf( outfile, "\tmovl %d(%%edx),%%eax\n", CONTEXTOFFSET(Eax) );
 650         fprintf( outfile, "\tmovl %d(%%edx),%%ebx\n", CONTEXTOFFSET(Ebx) );
 651         fprintf( outfile, "\tmovl %d(%%edx),%%ecx\n", CONTEXTOFFSET(Ecx) );
 652         fprintf( outfile, "\tmovl %d(%%edx),%%edx\n", CONTEXTOFFSET(Edx) );
 653
 654         /* Get the 16-bit ds */
 655         fprintf( outfile, "\tpopw %%ds\n" );
 656     }
 657     else  /* not a register function */
 658     {
 659         /* Push the called routine address */
 660         fprintf( outfile, "\tpushl %d(%%edx)\n", STACK32OFFSET(target) );
 661
 662         /* Set %fs to the value saved by the last CallFrom16 */
 663         fprintf( outfile, "\tmovw %d(%%ebp),%%ax\n", STACK16OFFSET(fs)-STACK16OFFSET(bp) );
 664         fprintf( outfile, "\tmovw %%ax,%%fs\n" );
 665
 666         /* Set %ds and %es (and %ax just in case) equal to %ss */
 667         fprintf( outfile, "\tmovw %%ss,%%ax\n" );
 668         fprintf( outfile, "\tmovw %%ax,%%ds\n" );
 669         fprintf( outfile, "\tmovw %%ax,%%es\n" );
 670     }
 671
 672     /* Jump to the called routine */
 673     fprintf( outfile, "\t.byte 0x66\n" );
 674     fprintf( outfile, "\tlret\n" );
 675 }
 676
 677
 678 /*******************************************************************
 679  *         BuildRet16Func
 680  *
 681  * Build the return code for 16-bit callbacks
 682  */
 683 static void BuildRet16Func( FILE *outfile )
 684 {
 685     /*
 686      *  Note: This must reside in the .data section to allow
 687      *        run-time relocation of the SYSLEVEL_Win16CurrentTeb symbol
 688      */
 689
 690     fprintf( outfile, "\n\t.type " PREFIX "CallTo16_Ret,@function\n" );
 691     fprintf( outfile, "\t.globl " PREFIX "CallTo16_Ret\n" );
 692     fprintf( outfile, PREFIX "CallTo16_Ret:\n" );
 693
 694     /* Save %esp into %esi */
 695     fprintf( outfile, "\tmovl %%esp,%%esi\n" );
 696
 697     /* Restore 32-bit segment registers */
 698
 699     fprintf( outfile, "\t.byte 0x2e\n\tmovl " PREFIX "CallTo16_DataSelector-" PREFIX "Call16_Ret_Start,%%edi\n" );
 700 #ifdef __svr4__
 701     fprintf( outfile, "\tdata16\n");
 702 #endif
 703     fprintf( outfile, "\tmovw %%di,%%ds\n" );
 704 #ifdef __svr4__
 705     fprintf( outfile, "\tdata16\n");
 706 #endif
 707     fprintf( outfile, "\tmovw %%di,%%es\n" );
 708
 709     fprintf( outfile, "\tmovw " PREFIX "SYSLEVEL_Win16CurrentTeb,%%fs\n" );
 710
 711     /* Restore the 32-bit stack */
 712
 713 #ifdef __svr4__
 714     fprintf( outfile, "\tdata16\n");
 715 #endif
 716     fprintf( outfile, "\tmovw %%di,%%ss\n" );
 717     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d),%%esp\n", STACKOFFSET );
 718     fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 719
 720     /* Return to caller */
 721
 722     fprintf( outfile, "\tlret\n" );
 723
 724     /* Declare the return address and data selector variables */
 725
 726     fprintf( outfile, "\n\t.align 4\n" );
 727     fprintf( outfile, "\t.globl " PREFIX "CallTo16_DataSelector\n" );
 728     fprintf( outfile, PREFIX "CallTo16_DataSelector:\t.long 0\n" );
 729     fprintf( outfile, "\t.globl " PREFIX "CallTo16_RetAddr\n" );
 730     fprintf( outfile, PREFIX "CallTo16_RetAddr:\t.long 0\n" );
 731 }
 732
 733
 734 /*******************************************************************
 735  *         BuildCallTo32CBClient
 736  *
 737  * Call a CBClient relay stub from 32-bit code (KERNEL.620).
 738  *
 739  * Since the relay stub is itself 32-bit, this should not be a problem;
 740  * unfortunately, the relay stubs are expected to switch back to a
 741  * 16-bit stack (and 16-bit code) after completion :-(
 742  *
 743  * This would conflict with our 16- vs. 32-bit stack handling, so
 744  * we simply switch *back* to our 32-bit stack before returning to
 745  * the caller ...
 746  *
 747  * The CBClient relay stub expects to be called with the following
 748  * 16-bit stack layout, and with ebp and ebx pointing into the 16-bit
 749  * stack at the designated places:
 750  *
 751  *    ...
 752  *  (ebp+14) original arguments to the callback routine
 753  *  (ebp+10) far return address to original caller
 754  *  (ebp+6)  Thunklet target address
 755  *  (ebp+2)  Thunklet relay ID code
 756  *  (ebp)    BP (saved by CBClientGlueSL)
 757  *  (ebp-2)  SI (saved by CBClientGlueSL)
 758  *  (ebp-4)  DI (saved by CBClientGlueSL)
 759  *  (ebp-6)  DS (saved by CBClientGlueSL)
 760  *
 761  *   ...     buffer space used by the 16-bit side glue for temp copies
 762  *
 763  *  (ebx+4)  far return address to 16-bit side glue code
 764  *  (ebx)    saved 16-bit ss:sp (pointing to ebx+4)
 765  *
 766  * The 32-bit side glue code accesses both the original arguments (via ebp)
 767  * and the temporary copies prepared by the 16-bit side glue (via ebx).
 768  * After completion, the stub will load ss:sp from the buffer at ebx
 769  * and perform a far return to 16-bit code.
 770  *
 771  * To trick the relay stub into returning to us, we replace the 16-bit
 772  * return address to the glue code by a cs:ip pair pointing to our
 773  * return entry point (the original return address is saved first).
 774  * Our return stub thus called will then reload the 32-bit ss:esp and
 775  * return to 32-bit code (by using and ss:esp value that we have also
 776  * pushed onto the 16-bit stack before and a cs:eip values found at
 777  * that position on the 32-bit stack).  The ss:esp to be restored is
 778  * found relative to the 16-bit stack pointer at:
 779  *
 780  *  (ebx-4)   ss  (flat)
 781  *  (ebx-8)   sp  (32-bit stack pointer)
 782  *
 783  * The second variant of this routine, CALL32_CBClientEx, which is used
 784  * to implement KERNEL.621, has to cope with yet another problem: Here,
 785  * the 32-bit side directly returns to the caller of the CBClient thunklet,
 786  * restoring registers saved by CBClientGlueSL and cleaning up the stack.
 787  * As we have to return to our 32-bit code first, we have to adapt the
 788  * layout of our temporary area so as to include values for the registers
 789  * that are to be restored, and later (in the implementation of KERNEL.621)
 790  * we *really* restore them. The return stub restores DS, DI, SI, and BP
 791  * from the stack, skips the next 8 bytes (CBClient relay code / target),
 792  * and then performs a lret NN, where NN is the number of arguments to be
 793  * removed. Thus, we prepare our temporary area as follows:
 794  *
 795  *     (ebx+22) 16-bit cs  (this segment)
 796  *     (ebx+20) 16-bit ip  ('16-bit' return entry point)
 797  *     (ebx+16) 32-bit ss  (flat)
 798  *     (ebx+12) 32-bit sp  (32-bit stack pointer)
 799  *     (ebx+10) 16-bit bp  (points to ebx+24)
 800  *     (ebx+8)  16-bit si  (ignored)
 801  *     (ebx+6)  16-bit di  (ignored)
 802  *     (ebx+4)  16-bit ds  (we actually use the flat DS here)
 803  *     (ebx+2)  16-bit ss  (16-bit stack segment)
 804  *     (ebx+0)  16-bit sp  (points to ebx+4)
 805  *
 806  * Note that we ensure that DS is not changed and remains the flat segment,
 807  * and the 32-bit stack pointer our own return stub needs fits just
 808  * perfectly into the 8 bytes that are skipped by the Windows stub.
 809  * One problem is that we have to determine the number of removed arguments,
 810  * as these have to be really removed in KERNEL.621. Thus, the BP value
 811  * that we place in the temporary area to be restored, contains the value
 812  * that SP would have if no arguments were removed. By comparing the actual
 813  * value of SP with this value in our return stub we can compute the number
 814  * of removed arguments. This is then returned to KERNEL.621.
 815  *
 816  * The stack layout of this function:
 817  * (ebp+20)  nArgs     pointer to variable receiving nr. of args (Ex only)
 818  * (ebp+16)  esi       pointer to caller's esi value
 819  * (ebp+12)  arg       ebp value to be set for relay stub
 820  * (ebp+8)   func      CBClient relay stub address
 821  * (ebp+4)   ret addr
 822  * (ebp)     ebp
 823  */
 824 static void BuildCallTo32CBClient( FILE *outfile, BOOL isEx )
 825 {
 826     char *name = isEx? "CBClientEx" : "CBClient";
 827     int size = isEx? 24 : 12;
 828
 829     /* Function header */
 830
 831     fprintf( outfile, "\n\t.align 4\n" );
 832 #ifdef USE_STABS
 833     fprintf( outfile, ".stabs \"CALL32_%s:F1\",36,0,0," PREFIX "CALL32_%s\n",
 834                       name, name );
 835 #endif
 836     fprintf( outfile, "\t.globl " PREFIX "CALL32_%s\n", name );
 837     fprintf( outfile, PREFIX "CALL32_%s:\n", name );
 838
 839     /* Entry code */
 840
 841     fprintf( outfile, "\tpushl %%ebp\n" );
 842     fprintf( outfile, "\tmovl %%esp,%%ebp\n" );
 843     fprintf( outfile, "\tpushl %%edi\n" );
 844     fprintf( outfile, "\tpushl %%esi\n" );
 845     fprintf( outfile, "\tpushl %%ebx\n" );
 846
 847     /* Get the 16-bit stack */
 848
 849     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d),%%ebx\n", STACKOFFSET);
 850
 851     /* Convert it to a flat address */
 852
 853     fprintf( outfile, "\tshldl $16,%%ebx,%%eax\n" );
 854     fprintf( outfile, "\tandl $0xfff8,%%eax\n" );
 855     fprintf( outfile, "\tmovl " PREFIX "ldt_copy(%%eax),%%esi\n" );
 856     fprintf( outfile, "\tmovw %%bx,%%ax\n" );
 857     fprintf( outfile, "\taddl %%eax,%%esi\n" );
 858
 859     /* Allocate temporary area (simulate STACK16_PUSH) */
 860
 861     fprintf( outfile, "\tpushf\n" );
 862     fprintf( outfile, "\tcld\n" );
 863     fprintf( outfile, "\tleal -%d(%%esi), %%edi\n", size );
 864     fprintf( outfile, "\tmovl $%d, %%ecx\n", sizeof(STACK16FRAME) );
 865     fprintf( outfile, "\trep\n\tmovsb\n" );
 866     fprintf( outfile, "\tpopf\n" );
 867
 868     fprintf( outfile, "\t.byte 0x64\n\tsubw $%d,(%d)\n", size, STACKOFFSET );
 869
 870     fprintf( outfile, "\tpushl %%edi\n" );  /* remember address */
 871
 872     /* Set up temporary area */
 873
 874     if ( !isEx )
 875     {
 876         fprintf( outfile, "\tleal 4(%%edi), %%edi\n" );
 877
 878         fprintf( outfile, "\tleal -8(%%esp), %%eax\n" );
 879         fprintf( outfile, "\tmovl %%eax, -8(%%edi)\n" );    /* 32-bit sp */
 880
 881         fprintf( outfile, "\tmovw %%ss, %%ax\n" );
 882         fprintf( outfile, "\tandl $0x0000ffff, %%eax\n" );
 883         fprintf( outfile, "\tmovl %%eax, -4(%%edi)\n" );    /* 32-bit ss */
 884
 885         fprintf( outfile, "\taddl $%d, %%ebx\n", sizeof(STACK16FRAME)-size+4 + 4 );
 886         fprintf( outfile, "\tmovl %%ebx, 0(%%edi)\n" );    /* 16-bit ss:sp */
 887
 888         fprintf( outfile, "\tmovl " PREFIX "CALL32_%s_RetAddr, %%eax\n", name );
 889         fprintf( outfile, "\tmovl %%eax, 4(%%edi)\n" );   /* overwrite return address */
 890     }
 891     else
 892     {
 893         fprintf( outfile, "\taddl $%d, %%ebx\n", sizeof(STACK16FRAME)-size+4 );
 894         fprintf( outfile, "\tmovl %%ebx, 0(%%edi)\n" );
 895
 896         fprintf( outfile, "\tmovw %%ds, %%ax\n" );
 897         fprintf( outfile, "\tmovw %%ax, 4(%%edi)\n" );
 898
 899         fprintf( outfile, "\taddl $20, %%ebx\n" );
 900         fprintf( outfile, "\tmovw %%bx, 10(%%edi)\n" );
 901
 902         fprintf( outfile, "\tleal -8(%%esp), %%eax\n" );
 903         fprintf( outfile, "\tmovl %%eax, 12(%%edi)\n" );
 904
 905         fprintf( outfile, "\tmovw %%ss, %%ax\n" );
 906         fprintf( outfile, "\tandl $0x0000ffff, %%eax\n" );
 907         fprintf( outfile, "\tmovl %%eax, 16(%%edi)\n" );
 908
 909         fprintf( outfile, "\tmovl " PREFIX "CALL32_%s_RetAddr, %%eax\n", name );
 910         fprintf( outfile, "\tmovl %%eax, 20(%%edi)\n" );
 911     }
 912
 913     /* Set up registers and call CBClient relay stub (simulating a far call) */
 914
 915     fprintf( outfile, "\tmovl 16(%%ebp), %%esi\n" );
 916     fprintf( outfile, "\tmovl (%%esi), %%esi\n" );
 917
 918     fprintf( outfile, "\tmovl %%edi, %%ebx\n" );
 919     fprintf( outfile, "\tmovl 8(%%ebp), %%eax\n" );
 920     fprintf( outfile, "\tmovl 12(%%ebp), %%ebp\n" );
 921
 922     fprintf( outfile, "\tpushl %%cs\n" );
 923     fprintf( outfile, "\tcall *%%eax\n" );
 924
 925     /* Return new esi value to caller */
 926
 927     fprintf( outfile, "\tmovl 32(%%esp), %%edi\n" );
 928     fprintf( outfile, "\tmovl %%esi, (%%edi)\n" );
 929
 930     /* Cleanup temporary area (simulate STACK16_POP) */
 931
 932     fprintf( outfile, "\tpop %%esi\n" );
 933
 934     fprintf( outfile, "\tpushf\n" );
 935     fprintf( outfile, "\tstd\n" );
 936     fprintf( outfile, "\tdec %%esi\n" );
 937     fprintf( outfile, "\tleal %d(%%esi), %%edi\n", size );
 938     fprintf( outfile, "\tmovl $%d, %%ecx\n", sizeof(STACK16FRAME) );
 939     fprintf( outfile, "\trep\n\tmovsb\n" );
 940     fprintf( outfile, "\tpopf\n" );
 941
 942     fprintf( outfile, "\t.byte 0x64\n\taddw $%d,(%d)\n", size, STACKOFFSET );
 943
 944     /* Return argument size to caller */
 945     if ( isEx )
 946     {
 947         fprintf( outfile, "\tmovl 32(%%esp), %%ebx\n" );
 948         fprintf( outfile, "\tmovl %%ebp, (%%ebx)\n" );
 949     }
 950
 951     /* Restore registers and return */
 952
 953     fprintf( outfile, "\tpopl %%ebx\n" );
 954     fprintf( outfile, "\tpopl %%esi\n" );
 955     fprintf( outfile, "\tpopl %%edi\n" );
 956     fprintf( outfile, "\tpopl %%ebp\n" );
 957     fprintf( outfile, "\tret\n" );
 958 }
 959
 960 static void BuildCallTo32CBClientRet( FILE *outfile, BOOL isEx )
 961 {
 962     char *name = isEx? "CBClientEx" : "CBClient";
 963
 964     /* '16-bit' return stub */
 965
 966     fprintf( outfile, "\n\t.globl " PREFIX "CALL32_%s_Ret\n", name );
 967     fprintf( outfile, PREFIX "CALL32_%s_Ret:\n", name );
 968
 969     if ( !isEx )
 970     {
 971         fprintf( outfile, "\tmovzwl %%sp, %%ebx\n" );
 972         fprintf( outfile, "\tlssl %%ss:-16(%%ebx), %%esp\n" );
 973     }
 974     else
 975     {
 976         fprintf( outfile, "\tmovzwl %%bp, %%ebx\n" );
 977         fprintf( outfile, "\tsubw %%bp, %%sp\n" );
 978         fprintf( outfile, "\tmovzwl %%sp, %%ebp\n" );
 979         fprintf( outfile, "\tlssl %%ss:-12(%%ebx), %%esp\n" );
 980     }
 981     fprintf( outfile, "\tlret\n" );
 982
 983     /* Declare the return address variable */
 984
 985     fprintf( outfile, "\n\t.globl " PREFIX "CALL32_%s_RetAddr\n", name );
 986     fprintf( outfile, PREFIX "CALL32_%s_RetAddr:\t.long 0\n", name );
 987 }
 988
 989
 990 /*******************************************************************
 991  *         BuildCallFrom32Regs
 992  *
 993  * Build a 32-bit-to-Wine call-back function for a 'register' function.
 994  * 'args' is the number of dword arguments.
 995  *
 996  * Stack layout:
 997  *   ...
 998  * (ebp+12)  first arg
 999  * (ebp+8)   ret addr to user code
1000  * (ebp+4)   ret addr to relay code
1001  * (ebp+0)   saved ebp
1002  * (ebp-128) buffer area to allow stack frame manipulation
1003  * (ebp-332) CONTEXT86 struct
1004  * (ebp-336) CONTEXT86 *argument
1005  *  ....     other arguments copied from (ebp+12)
1006  *
1007  * The entry point routine is called with a CONTEXT* extra argument,
1008  * following the normal args. In this context structure, EIP_reg
1009  * contains the return address to user code, and ESP_reg the stack
1010  * pointer on return (with the return address and arguments already
1011  * removed).
1012  */
1013 static void BuildCallFrom32Regs( FILE *outfile )
1014 {
1015     static const int STACK_SPACE = 128 + sizeof(CONTEXT86);
1016
1017     /* Function header */
1018
1019     fprintf( outfile, "\n\t.align 4\n" );
1020 #ifdef USE_STABS
1021     fprintf( outfile, ".stabs \"CALL32_Regs:F1\",36,0,0," PREFIX "CALL32_Regs\n" );
1022 #endif
1023     fprintf( outfile, "\t.globl " PREFIX "CALL32_Regs\n" );
1024     fprintf( outfile, PREFIX "CALL32_Regs:\n" );
1025
1026     /* Allocate some buffer space on the stack */
1027
1028     fprintf( outfile, "\tpushl %%ebp\n" );
1029     fprintf( outfile, "\tmovl %%esp,%%ebp\n ");
1030     fprintf( outfile, "\tleal -%d(%%esp), %%esp\n", STACK_SPACE );
1031
1032     /* Build the context structure */
1033
1034     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Eax) - STACK_SPACE );
1035     fprintf( outfile, "\tpushfl\n" );
1036     fprintf( outfile, "\tpopl %%eax\n" );
1037     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(EFlags) - STACK_SPACE );
1038     fprintf( outfile, "\tmovl 0(%%ebp),%%eax\n" );
1039     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Ebp) - STACK_SPACE );
1040     fprintf( outfile, "\tmovl %%ebx,%d(%%ebp)\n", CONTEXTOFFSET(Ebx) - STACK_SPACE );
1041     fprintf( outfile, "\tmovl %%ecx,%d(%%ebp)\n", CONTEXTOFFSET(Ecx) - STACK_SPACE );
1042     fprintf( outfile, "\tmovl %%edx,%d(%%ebp)\n", CONTEXTOFFSET(Edx) - STACK_SPACE );
1043     fprintf( outfile, "\tmovl %%esi,%d(%%ebp)\n", CONTEXTOFFSET(Esi) - STACK_SPACE );
1044     fprintf( outfile, "\tmovl %%edi,%d(%%ebp)\n", CONTEXTOFFSET(Edi) - STACK_SPACE );
1045
1046     fprintf( outfile, "\txorl %%eax,%%eax\n" );
1047     fprintf( outfile, "\tmovw %%cs,%%ax\n" );
1048     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegCs) - STACK_SPACE );
1049     fprintf( outfile, "\tmovw %%es,%%ax\n" );
1050     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegEs) - STACK_SPACE );
1051     fprintf( outfile, "\tmovw %%fs,%%ax\n" );
1052     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegFs) - STACK_SPACE );
1053     fprintf( outfile, "\tmovw %%gs,%%ax\n" );
1054     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegGs) - STACK_SPACE );
1055     fprintf( outfile, "\tmovw %%ss,%%ax\n" );
1056     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegSs) - STACK_SPACE );
1057     fprintf( outfile, "\tmovw %%ds,%%ax\n" );
1058     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegDs) - STACK_SPACE );
1059     fprintf( outfile, "\tmovw %%ax,%%es\n" );  /* set %es equal to %ds just in case */
1060
1061     fprintf( outfile, "\tmovl $0x%x,%%eax\n", CONTEXT86_FULL );
1062     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(ContextFlags) - STACK_SPACE );
1063
1064     fprintf( outfile, "\tmovl 8(%%ebp),%%eax\n" ); /* Get %eip at time of call */
1065     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Eip) - STACK_SPACE );
1066
1067     /* Transfer the arguments */
1068
1069     fprintf( outfile, "\tmovl 4(%%ebp),%%ebx\n" );   /* get relay code addr */
1070     fprintf( outfile, "\tpushl %%esp\n" );           /* push ptr to context struct */
1071     fprintf( outfile, "\tmovzbl 4(%%ebx),%%ecx\n" ); /* fetch number of args to copy */
1072     fprintf( outfile, "\tjecxz 1f\n" );
1073     fprintf( outfile, "\tsubl %%ecx,%%esp\n" );
1074     fprintf( outfile, "\tleal 12(%%ebp),%%esi\n" );  /* get %esp at time of call */
1075     fprintf( outfile, "\tmovl %%esp,%%edi\n" );
1076     fprintf( outfile, "\tshrl $2,%%ecx\n" );
1077     fprintf( outfile, "\tcld\n" );
1078     fprintf( outfile, "\trep\n\tmovsl\n" );  /* copy args */
1079
1080     fprintf( outfile, "1:\tmovzbl 5(%%ebx),%%eax\n" ); /* fetch number of args to remove */
1081     fprintf( outfile, "\tleal 12(%%ebp,%%eax),%%eax\n" );
1082     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Esp) - STACK_SPACE );
1083
1084     /* Call the entry point */
1085
1086     fprintf( outfile, "\tcall *0(%%ebx)\n" );
1087
1088     /* Store %eip and %ebp onto the new stack */
1089
1090     fprintf( outfile, "\tmovl %d(%%ebp),%%edx\n", CONTEXTOFFSET(Esp) - STACK_SPACE );
1091     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(Eip) - STACK_SPACE );
1092     fprintf( outfile, "\tmovl %%eax,-4(%%edx)\n" );
1093     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(Ebp) - STACK_SPACE );
1094     fprintf( outfile, "\tmovl %%eax,-8(%%edx)\n" );
1095
1096     /* Restore the context structure */
1097
1098     /* Note: we don't bother to restore %cs, %ds and %ss
1099      *       changing them in 32-bit code is a recipe for disaster anyway
1100      */
1101     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(SegEs) - STACK_SPACE );
1102     fprintf( outfile, "\tmovw %%ax,%%es\n" );
1103     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(SegFs) - STACK_SPACE );
1104     fprintf( outfile, "\tmovw %%ax,%%fs\n" );
1105     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(SegGs) - STACK_SPACE );
1106     fprintf( outfile, "\tmovw %%ax,%%gs\n" );
1107
1108     fprintf( outfile, "\tmovl %d(%%ebp),%%edi\n", CONTEXTOFFSET(Edi) - STACK_SPACE );
1109     fprintf( outfile, "\tmovl %d(%%ebp),%%esi\n", CONTEXTOFFSET(Esi) - STACK_SPACE );
1110     fprintf( outfile, "\tmovl %d(%%ebp),%%edx\n", CONTEXTOFFSET(Edx) - STACK_SPACE );
1111     fprintf( outfile, "\tmovl %d(%%ebp),%%ecx\n", CONTEXTOFFSET(Ecx) - STACK_SPACE );
1112     fprintf( outfile, "\tmovl %d(%%ebp),%%ebx\n", CONTEXTOFFSET(Ebx) - STACK_SPACE );
1113
1114     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(EFlags) - STACK_SPACE );
1115     fprintf( outfile, "\tpushl %%eax\n" );
1116     fprintf( outfile, "\tpopfl\n" );
1117     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(Eax) - STACK_SPACE );
1118
1119     fprintf( outfile, "\tmovl %d(%%ebp),%%ebp\n", CONTEXTOFFSET(Esp) - STACK_SPACE );
1120     fprintf( outfile, "\tleal -8(%%ebp),%%esp\n" );
1121     fprintf( outfile, "\tpopl %%ebp\n" );
1122     fprintf( outfile, "\tret\n" );
1123 }
1124
1125
1126 /*******************************************************************
1127  *         BuildRelays
1128  *
1129  * Build all the relay callbacks
1130  */
1131 void BuildRelays( FILE *outfile )
1132 {
1133     /* File header */
1134
1135     fprintf( outfile, "/* File generated automatically. Do not edit! */\n\n" );
1136     fprintf( outfile, "\t.text\n" );
1137
1138 #ifdef __i386__
1139
1140 #ifdef USE_STABS
1141     if (output_file_name)
1142     {
1143         char buffer[1024];
1144         getcwd(buffer, sizeof(buffer));
1145         fprintf( outfile, "\t.file\t\"%s\"\n", output_file_name );
1146
1147         /*
1148          * The stabs help the internal debugger as they are an indication that it
1149          * is sensible to step into a thunk/trampoline.
1150          */
1151         fprintf( outfile, ".stabs \"%s/\",100,0,0,Code_Start\n", buffer);
1152         fprintf( outfile, ".stabs \"%s\",100,0,0,Code_Start\n", output_file_name );
1153         fprintf( outfile, "Code_Start:\n\n" );
1154     }
1155 #endif
1156     fprintf( outfile, PREFIX"Call16_Start:\n" );
1157     fprintf( outfile, "\t.globl "PREFIX"Call16_Start\n" );
1158     fprintf( outfile, "\t.byte 0\n\n" );
1159
1160     /* Standard CallFrom16 routine (WORD return) */
1161     BuildCallFrom16Core( outfile, FALSE, FALSE, TRUE );
1162
1163     /* Standard CallFrom16 routine (DWORD return) */
1164     BuildCallFrom16Core( outfile, FALSE, FALSE, FALSE );
1165
1166     /* Register CallFrom16 routine */
1167     BuildCallFrom16Core( outfile, TRUE, FALSE, FALSE );
1168
1169     /* C16ThkSL CallFrom16 routine */
1170     BuildCallFrom16Core( outfile, FALSE, TRUE, FALSE );
1171
1172     /* Standard CallTo16 routine (WORD return) */
1173     BuildCallTo16Core( outfile, TRUE, FALSE );
1174
1175     /* Standard CallTo16 routine (DWORD return) */
1176     BuildCallTo16Core( outfile, FALSE, FALSE );
1177
1178     /* Register CallTo16 routine (16:16 retf) */
1179     BuildCallTo16Core( outfile, FALSE, 1 );
1180
1181     /* Register CallTo16 routine (16:32 retf) */
1182     BuildCallTo16Core( outfile, FALSE, 2 );
1183
1184     /* CBClientThunkSL routine */
1185     BuildCallTo32CBClient( outfile, FALSE );
1186
1187     /* CBClientThunkSLEx routine */
1188     BuildCallTo32CBClient( outfile, TRUE  );
1189
1190     /* 32-bit register entry point */
1191     BuildCallFrom32Regs( outfile );
1192
1193     fprintf( outfile, PREFIX"Call16_End:\n" );
1194     fprintf( outfile, "\t.globl "PREFIX"Call16_End\n" );
1195
1196 #ifdef USE_STABS
1197     fprintf( outfile, "\t.stabs \"\",100,0,0,.Letext\n");
1198     fprintf( outfile, ".Letext:\n");
1199 #endif
1200
1201     /* The whole Call16_Ret segment must lie within the .data section */
1202     fprintf( outfile, "\n\t.data\n" );
1203     fprintf( outfile, "\t.globl " PREFIX "Call16_Ret_Start\n" );
1204     fprintf( outfile, PREFIX "Call16_Ret_Start:\n" );
1205
1206     /* Standard CallTo16 return stub */
1207     BuildRet16Func( outfile );
1208
1209     /* CBClientThunkSL return stub */
1210     BuildCallTo32CBClientRet( outfile, FALSE );
1211
1212     /* CBClientThunkSLEx return stub */
1213     BuildCallTo32CBClientRet( outfile, TRUE  );
1214
1215     /* End of Call16_Ret segment */
1216     fprintf( outfile, "\n\t.globl " PREFIX "Call16_Ret_End\n" );
1217     fprintf( outfile, PREFIX "Call16_Ret_End:\n" );
1218
1219 #else  /* __i386__ */
1220
1221     fprintf( outfile, PREFIX"Call16_Start:\n" );
1222     fprintf( outfile, "\t.globl "PREFIX"Call16_Start\n" );
1223     fprintf( outfile, "\t.byte 0\n\n" );
1224     fprintf( outfile, PREFIX"Call16_End:\n" );
1225     fprintf( outfile, "\t.globl "PREFIX"Call16_End\n" );
1226
1227     fprintf( outfile, "\t.globl " PREFIX "Call16_Ret_Start\n" );
1228     fprintf( outfile, PREFIX "Call16_Ret_Start:\n" );
1229     fprintf( outfile, "\t.byte 0\n\n" );
1230     fprintf( outfile, "\n\t.globl " PREFIX "Call16_Ret_End\n" );
1231     fprintf( outfile, PREFIX "Call16_Ret_End:\n" );
1232
1233 #endif  /* __i386__ */
1234 }