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 "config.h"
  12
  13 #include <ctype.h>
  14 #include <unistd.h>
  15
  16 #include "winnt.h"
  17 #include "thread.h"
  18 #include "stackframe.h"
  19
  20 #include "build.h"
  21
  22 #ifdef __i386__
  23
  24 static void function_header( FILE *outfile, const char *name )
  25 {
  26     fprintf( outfile, "\n\t.align %d\n", get_alignment(4) );
  27 #ifdef USE_STABS
  28     fprintf( outfile, "\t.stabs \"%s:F1\",36,0,0," PREFIX "%s\n", name, name);
  29 #endif
  30 #ifdef NEED_TYPE_IN_DEF
  31     fprintf( outfile, "\t.def " PREFIX "%s; .scl 2; .type 32; .endef\n", name );
  32 #else
  33     fprintf( outfile, "\t.type " PREFIX "%s,@function\n", name );
  34 #endif
  35     fprintf( outfile, "\t.globl " PREFIX "%s\n", name );
  36     fprintf( outfile, PREFIX "%s:\n", name );
  37 }
  38
  39
  40 /*******************************************************************
  41  *         BuildCallFrom16Core
  42  *
  43  * This routine builds the core routines used in 16->32 thunks:
  44  * CallFrom16Word, CallFrom16Long, CallFrom16Register, and CallFrom16Thunk.
  45  *
  46  * These routines are intended to be called via a far call (with 32-bit
  47  * operand size) from 16-bit code.  The 16-bit code stub must push %bp,
  48  * the 32-bit entry point to be called, and the argument conversion
  49  * routine to be used (see stack layout below).
  50  *
  51  * The core routine completes the STACK16FRAME on the 16-bit stack and
  52  * switches to the 32-bit stack.  Then, the argument conversion routine
  53  * is called; it gets passed the 32-bit entry point and a pointer to the
  54  * 16-bit arguments (on the 16-bit stack) as parameters. (You can either
  55  * use conversion routines automatically generated by BuildCallFrom16,
  56  * or write your own for special purposes.)
  57  *
  58  * The conversion routine must call the 32-bit entry point, passing it
  59  * the converted arguments, and return its return value to the core.
  60  * After the conversion routine has returned, the core switches back
  61  * to the 16-bit stack, converts the return value to the DX:AX format
  62  * (CallFrom16Long), and returns to the 16-bit call stub.  All parameters,
  63  * including %bp, are popped off the stack.
  64  *
  65  * The 16-bit call stub now returns to the caller, popping the 16-bit
  66  * arguments if necessary (pascal calling convention).
  67  *
  68  * In the case of a 'register' function, CallFrom16Register fills a
  69  * CONTEXT86 structure with the values all registers had at the point
  70  * the first instruction of the 16-bit call stub was about to be
  71  * executed.  A pointer to this CONTEXT86 is passed as third parameter
  72  * to the argument conversion routine, which typically passes it on
  73  * to the called 32-bit entry point.
  74  *
  75  * CallFrom16Thunk is a special variant used by the implementation of
  76  * the Win95 16->32 thunk functions C16ThkSL and C16ThkSL01 and is
  77  * implemented as follows:
  78  * On entry, the EBX register is set up to contain a flat pointer to the
  79  * 16-bit stack such that EBX+22 points to the first argument.
  80  * Then, the entry point is called, while EBP is set up to point
  81  * to the return address (on the 32-bit stack).
  82  * The called function returns with CX set to the number of bytes
  83  * to be popped of the caller's stack.
  84  *
  85  * Stack layout upon entry to the core routine (STACK16FRAME):
  86  *  ...           ...
  87  * (sp+24) word   first 16-bit arg
  88  * (sp+22) word   cs
  89  * (sp+20) word   ip
  90  * (sp+18) word   bp
  91  * (sp+14) long   32-bit entry point (reused for Win16 mutex recursion count)
  92  * (sp+12) word   ip of actual entry point (necessary for relay debugging)
  93  * (sp+8)  long   relay (argument conversion) function entry point
  94  * (sp+4)  long   cs of 16-bit entry point
  95  * (sp)    long   ip of 16-bit entry point
  96  *
  97  * Added on the stack:
  98  * (sp-2)  word   saved gs
  99  * (sp-4)  word   saved fs
 100  * (sp-6)  word   saved es
 101  * (sp-8)  word   saved ds
 102  * (sp-12) long   saved ebp
 103  * (sp-16) long   saved ecx
 104  * (sp-20) long   saved edx
 105  * (sp-24) long   saved previous stack
 106  */
 107 static void BuildCallFrom16Core( FILE *outfile, int reg_func, int thunk, int short_ret )
 108 {
 109     char *name = thunk? "thunk" : reg_func? "regs" : short_ret? "word" : "long";
 110
 111     /* Function header */
 112     if (thunk) function_header( outfile, "__wine_call_from_16_thunk" );
 113     else if (reg_func) function_header( outfile, "__wine_call_from_16_regs" );
 114     else if (short_ret) function_header( outfile, "__wine_call_from_16_word" );
 115     else function_header( outfile, "__wine_call_from_16_long" );
 116
 117     /* Create STACK16FRAME (except STACK32FRAME link) */
 118     fprintf( outfile, "\tpushw %%gs\n" );
 119     fprintf( outfile, "\tpushw %%fs\n" );
 120     fprintf( outfile, "\tpushw %%es\n" );
 121     fprintf( outfile, "\tpushw %%ds\n" );
 122     fprintf( outfile, "\tpushl %%ebp\n" );
 123     fprintf( outfile, "\tpushl %%ecx\n" );
 124     fprintf( outfile, "\tpushl %%edx\n" );
 125
 126     /* Save original EFlags register */
 127     fprintf( outfile, "\tpushfl\n" );
 128
 129     if ( UsePIC )
 130     {
 131         /* Get Global Offset Table into %ecx */
 132         fprintf( outfile, "\tcall .L__wine_call_from_16_%s.getgot1\n", name );
 133         fprintf( outfile, ".L__wine_call_from_16_%s.getgot1:\n", name );
 134         fprintf( outfile, "\tpopl %%ecx\n" );
 135         fprintf( outfile, "\taddl $_GLOBAL_OFFSET_TABLE_+[.-.L__wine_call_from_16_%s.getgot1], %%ecx\n", name );
 136     }
 137
 138     if (UsePIC)
 139     {
 140         fprintf( outfile, "\t.byte 0x2e\n\tmovl " PREFIX "CallTo16_DataSelector@GOT(%%ecx), %%edx\n" );
 141         fprintf( outfile, "\t.byte 0x2e\n\tmovl (%%edx), %%edx\n" );
 142     }
 143     else
 144         fprintf( outfile, "\t.byte 0x2e\n\tmovl " PREFIX "CallTo16_DataSelector,%%edx\n" );
 145
 146     /* Load 32-bit segment registers */
 147 #ifdef __svr4__
 148     fprintf( outfile, "\tdata16\n");
 149 #endif
 150     fprintf( outfile, "\tmovw %%dx, %%ds\n" );
 151 #ifdef __svr4__
 152     fprintf( outfile, "\tdata16\n");
 153 #endif
 154     fprintf( outfile, "\tmovw %%dx, %%es\n" );
 155
 156     if ( UsePIC )
 157     {
 158         fprintf( outfile, "\tmovl " PREFIX "SYSLEVEL_Win16CurrentTeb@GOT(%%ecx), %%edx\n" );
 159         fprintf( outfile, "\tmovw (%%edx), %%fs\n" );
 160     }
 161     else
 162         fprintf( outfile, "\tmovw " PREFIX "SYSLEVEL_Win16CurrentTeb, %%fs\n" );
 163
 164     /* Get address of wine_ldt_copy array into %ecx */
 165     if ( UsePIC )
 166         fprintf( outfile, "\tmovl " PREFIX "wine_ldt_copy@GOT(%%ecx), %%ecx\n" );
 167     else
 168         fprintf( outfile, "\tmovl $" PREFIX "wine_ldt_copy, %%ecx\n" );
 169
 170     /* Translate STACK16FRAME base to flat offset in %edx */
 171     fprintf( outfile, "\tmovw %%ss, %%dx\n" );
 172     fprintf( outfile, "\tandl $0xfff8, %%edx\n" );
 173     fprintf( outfile, "\tshrl $1, %%edx\n" );
 174     fprintf( outfile, "\tmovl (%%ecx,%%edx), %%edx\n" );
 175     fprintf( outfile, "\tmovzwl %%sp, %%ebp\n" );
 176     fprintf( outfile, "\tleal (%%ebp,%%edx), %%edx\n" );
 177
 178     /* Get saved flags into %ecx */
 179     fprintf( outfile, "\tpopl %%ecx\n" );
 180
 181     /* Get the 32-bit stack pointer from the TEB and complete STACK16FRAME */
 182     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d), %%ebp\n", STACKOFFSET );
 183     fprintf( outfile, "\tpushl %%ebp\n" );
 184
 185     /* Switch stacks */
 186 #ifdef __svr4__
 187     fprintf( outfile,"\tdata16\n");
 188 #endif
 189     fprintf( outfile, "\t.byte 0x64\n\tmovw %%ss, (%d)\n", STACKOFFSET + 2 );
 190     fprintf( outfile, "\t.byte 0x64\n\tmovw %%sp, (%d)\n", STACKOFFSET );
 191     fprintf( outfile, "\tpushl %%ds\n" );
 192     fprintf( outfile, "\tpopl %%ss\n" );
 193     fprintf( outfile, "\tmovl %%ebp, %%esp\n" );
 194     fprintf( outfile, "\taddl $%d, %%ebp\n", STRUCTOFFSET(STACK32FRAME, ebp) );
 195
 196
 197     /* At this point:
 198        STACK16FRAME is completely set up
 199        DS, ES, SS: flat data segment
 200        FS: current TEB
 201        ESP: points to last STACK32FRAME
 202        EBP: points to ebp member of last STACK32FRAME
 203        EDX: points to current STACK16FRAME
 204        ECX: contains saved flags
 205        all other registers: unchanged */
 206
 207     /* Special case: C16ThkSL stub */
 208     if ( thunk )
 209     {
 210         /* Set up registers as expected and call thunk */
 211         fprintf( outfile, "\tleal %d(%%edx), %%ebx\n", sizeof(STACK16FRAME)-22 );
 212         fprintf( outfile, "\tleal -4(%%esp), %%ebp\n" );
 213
 214         fprintf( outfile, "\tcall *%d(%%edx)\n", STACK16OFFSET(entry_point) );
 215
 216         /* Switch stack back */
 217         fprintf( outfile, "\t.byte 0x64\n\tmovw (%d), %%ss\n", STACKOFFSET+2 );
 218         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%esp\n", STACKOFFSET );
 219         fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 220
 221         /* Restore registers and return directly to caller */
 222         fprintf( outfile, "\taddl $8, %%esp\n" );
 223         fprintf( outfile, "\tpopl %%ebp\n" );
 224         fprintf( outfile, "\tpopw %%ds\n" );
 225         fprintf( outfile, "\tpopw %%es\n" );
 226         fprintf( outfile, "\tpopw %%fs\n" );
 227         fprintf( outfile, "\tpopw %%gs\n" );
 228         fprintf( outfile, "\taddl $20, %%esp\n" );
 229
 230         fprintf( outfile, "\txorb %%ch, %%ch\n" );
 231         fprintf( outfile, "\tpopl %%ebx\n" );
 232         fprintf( outfile, "\taddw %%cx, %%sp\n" );
 233         fprintf( outfile, "\tpush %%ebx\n" );
 234
 235         fprintf( outfile, "\t.byte 0x66\n" );
 236         fprintf( outfile, "\tlret\n" );
 237
 238         return;
 239     }
 240
 241
 242     /* Build register CONTEXT */
 243     if ( reg_func )
 244     {
 245         fprintf( outfile, "\tsubl $%d, %%esp\n", sizeof(CONTEXT86) );
 246
 247         fprintf( outfile, "\tmovl %%ecx, %d(%%esp)\n", CONTEXTOFFSET(EFlags) );
 248
 249         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Eax) );
 250         fprintf( outfile, "\tmovl %%ebx, %d(%%esp)\n", CONTEXTOFFSET(Ebx) );
 251         fprintf( outfile, "\tmovl %%esi, %d(%%esp)\n", CONTEXTOFFSET(Esi) );
 252         fprintf( outfile, "\tmovl %%edi, %d(%%esp)\n", CONTEXTOFFSET(Edi) );
 253
 254         fprintf( outfile, "\tmovl %d(%%edx), %%eax\n", STACK16OFFSET(ebp) );
 255         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Ebp) );
 256         fprintf( outfile, "\tmovl %d(%%edx), %%eax\n", STACK16OFFSET(ecx) );
 257         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Ecx) );
 258         fprintf( outfile, "\tmovl %d(%%edx), %%eax\n", STACK16OFFSET(edx) );
 259         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Edx) );
 260
 261         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(ds) );
 262         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegDs) );
 263         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(es) );
 264         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegEs) );
 265         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(fs) );
 266         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegFs) );
 267         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(gs) );
 268         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegGs) );
 269
 270         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(cs) );
 271         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegCs) );
 272         fprintf( outfile, "\tmovzwl %d(%%edx), %%eax\n", STACK16OFFSET(ip) );
 273         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Eip) );
 274
 275         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%eax\n", STACKOFFSET+2 );
 276         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(SegSs) );
 277         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%eax\n", STACKOFFSET );
 278         fprintf( outfile, "\taddl $%d, %%eax\n", STACK16OFFSET(ip) );
 279         fprintf( outfile, "\tmovl %%eax, %d(%%esp)\n", CONTEXTOFFSET(Esp) );
 280 #if 0
 281         fprintf( outfile, "\tfsave %d(%%esp)\n", CONTEXTOFFSET(FloatSave) );
 282 #endif
 283
 284         /* Push address of CONTEXT86 structure -- popped by the relay routine */
 285         fprintf( outfile, "\tpushl %%esp\n" );
 286     }
 287
 288
 289     /* Print debug info before call */
 290     if ( debugging )
 291     {
 292         if ( UsePIC )
 293         {
 294             fprintf( outfile, "\tpushl %%ebx\n" );
 295
 296             /* Get Global Offset Table into %ebx (for PLT call) */
 297             fprintf( outfile, "\tcall .L__wine_call_from_16_%s.getgot2\n", name );
 298             fprintf( outfile, ".L__wine_call_from_16_%s.getgot2:\n", name );
 299             fprintf( outfile, "\tpopl %%ebx\n" );
 300             fprintf( outfile, "\taddl $_GLOBAL_OFFSET_TABLE_+[.-.L__wine_call_from_16_%s.getgot2], %%ebx\n", name );
 301         }
 302
 303         fprintf( outfile, "\tpushl %%edx\n" );
 304         if ( reg_func )
 305             fprintf( outfile, "\tleal -%d(%%ebp), %%eax\n\tpushl %%eax\n",
 306                               sizeof(CONTEXT) + STRUCTOFFSET(STACK32FRAME, ebp) );
 307         else
 308             fprintf( outfile, "\tpushl $0\n" );
 309
 310         if ( UsePIC )
 311             fprintf( outfile, "\tcall " PREFIX "RELAY_DebugCallFrom16@PLT\n ");
 312         else
 313             fprintf( outfile, "\tcall " PREFIX "RELAY_DebugCallFrom16\n ");
 314
 315         fprintf( outfile, "\tpopl %%edx\n" );
 316         fprintf( outfile, "\tpopl %%edx\n" );
 317
 318         if ( UsePIC )
 319             fprintf( outfile, "\tpopl %%ebx\n" );
 320     }
 321
 322     /* Call relay routine (which will call the API entry point) */
 323     fprintf( outfile, "\tleal %d(%%edx), %%eax\n", sizeof(STACK16FRAME) );
 324     fprintf( outfile, "\tpushl %%eax\n" );
 325     fprintf( outfile, "\tpushl %d(%%edx)\n", STACK16OFFSET(entry_point) );
 326     fprintf( outfile, "\tcall *%d(%%edx)\n", STACK16OFFSET(relay) );
 327
 328     /* Print debug info after call */
 329     if ( debugging )
 330     {
 331         if ( UsePIC )
 332         {
 333             fprintf( outfile, "\tpushl %%ebx\n" );
 334
 335             /* Get Global Offset Table into %ebx (for PLT call) */
 336             fprintf( outfile, "\tcall .L__wine_call_from_16_%s.getgot3\n", name );
 337             fprintf( outfile, ".L__wine_call_from_16_%s.getgot3:\n", name );
 338             fprintf( outfile, "\tpopl %%ebx\n" );
 339             fprintf( outfile, "\taddl $_GLOBAL_OFFSET_TABLE_+[.-.L__wine_call_from_16_%s.getgot3], %%ebx\n", name );
 340         }
 341
 342         fprintf( outfile, "\tpushl %%eax\n" );
 343         if ( reg_func )
 344             fprintf( outfile, "\tleal -%d(%%ebp), %%eax\n\tpushl %%eax\n",
 345                               sizeof(CONTEXT) + STRUCTOFFSET(STACK32FRAME, ebp) );
 346         else
 347             fprintf( outfile, "\tpushl $0\n" );
 348
 349         if ( UsePIC )
 350             fprintf( outfile, "\tcall " PREFIX "RELAY_DebugCallFrom16Ret@PLT\n ");
 351         else
 352             fprintf( outfile, "\tcall " PREFIX "RELAY_DebugCallFrom16Ret\n ");
 353
 354         fprintf( outfile, "\tpopl %%eax\n" );
 355         fprintf( outfile, "\tpopl %%eax\n" );
 356
 357         if ( UsePIC )
 358             fprintf( outfile, "\tpopl %%ebx\n" );
 359     }
 360
 361
 362     if ( reg_func )
 363     {
 364         fprintf( outfile, "\tmovl %%esp, %%ebx\n" );
 365
 366         /* Switch stack back */
 367         fprintf( outfile, "\t.byte 0x64\n\tmovw (%d), %%ss\n", STACKOFFSET+2 );
 368         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%esp\n", STACKOFFSET );
 369         fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 370
 371         /* Get return address to CallFrom16 stub */
 372         fprintf( outfile, "\taddw $%d, %%sp\n", STACK16OFFSET(callfrom_ip)-4 );
 373         fprintf( outfile, "\tpopl %%eax\n" );
 374         fprintf( outfile, "\tpopl %%edx\n" );
 375
 376         /* Restore all registers from CONTEXT */
 377         fprintf( outfile, "\tmovw %d(%%ebx), %%ss\n", CONTEXTOFFSET(SegSs) );
 378         fprintf( outfile, "\tmovl %d(%%ebx), %%esp\n", CONTEXTOFFSET(Esp) );
 379         fprintf( outfile, "\taddl $4, %%esp\n" );  /* room for final return address */
 380
 381         fprintf( outfile, "\tpushw %d(%%ebx)\n", CONTEXTOFFSET(SegCs) );
 382         fprintf( outfile, "\tpushw %d(%%ebx)\n", CONTEXTOFFSET(Eip) );
 383         fprintf( outfile, "\tpushl %%edx\n" );
 384         fprintf( outfile, "\tpushl %%eax\n" );
 385         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(EFlags) );
 386         fprintf( outfile, "\tpushl %d(%%ebx)\n", CONTEXTOFFSET(SegDs) );
 387
 388         fprintf( outfile, "\tmovw %d(%%ebx), %%es\n", CONTEXTOFFSET(SegEs) );
 389         fprintf( outfile, "\tmovw %d(%%ebx), %%fs\n", CONTEXTOFFSET(SegFs) );
 390         fprintf( outfile, "\tmovw %d(%%ebx), %%gs\n", CONTEXTOFFSET(SegGs) );
 391
 392         fprintf( outfile, "\tmovl %d(%%ebx), %%ebp\n", CONTEXTOFFSET(Ebp) );
 393         fprintf( outfile, "\tmovl %d(%%ebx), %%esi\n", CONTEXTOFFSET(Esi) );
 394         fprintf( outfile, "\tmovl %d(%%ebx), %%edi\n", CONTEXTOFFSET(Edi) );
 395         fprintf( outfile, "\tmovl %d(%%ebx), %%eax\n", CONTEXTOFFSET(Eax) );
 396         fprintf( outfile, "\tmovl %d(%%ebx), %%edx\n", CONTEXTOFFSET(Edx) );
 397         fprintf( outfile, "\tmovl %d(%%ebx), %%ecx\n", CONTEXTOFFSET(Ecx) );
 398         fprintf( outfile, "\tmovl %d(%%ebx), %%ebx\n", CONTEXTOFFSET(Ebx) );
 399
 400         fprintf( outfile, "\tpopl %%ds\n" );
 401         fprintf( outfile, "\tpopfl\n" );
 402         fprintf( outfile, "\tlret\n" );
 403     }
 404     else
 405     {
 406         /* Switch stack back */
 407         fprintf( outfile, "\t.byte 0x64\n\tmovw (%d), %%ss\n", STACKOFFSET+2 );
 408         fprintf( outfile, "\t.byte 0x64\n\tmovzwl (%d), %%esp\n", STACKOFFSET );
 409         fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 410
 411         /* Restore registers */
 412         fprintf( outfile, "\tpopl %%edx\n" );
 413         fprintf( outfile, "\tpopl %%ecx\n" );
 414         fprintf( outfile, "\tpopl %%ebp\n" );
 415         fprintf( outfile, "\tpopw %%ds\n" );
 416         fprintf( outfile, "\tpopw %%es\n" );
 417         fprintf( outfile, "\tpopw %%fs\n" );
 418         fprintf( outfile, "\tpopw %%gs\n" );
 419
 420         /* Prepare return value and set flags accordingly */
 421         if ( !short_ret )
 422             fprintf( outfile, "\tshldl $16, %%eax, %%edx\n" );
 423         fprintf( outfile, "\torl %%eax, %%eax\n" );
 424
 425         /* Return to return stub which will return to caller */
 426         fprintf( outfile, "\tlret $12\n" );
 427     }
 428 }
 429
 430
 431 /*******************************************************************
 432  *         BuildCallTo16Core
 433  *
 434  * This routine builds the core routines used in 32->16 thunks:
 435  *
 436  *   extern void WINAPI wine_call_to_16_word( SEGPTR target, int nb_args );
 437  *   extern void WINAPI wine_call_to_16_long( SEGPTR target, int nb_args );
 438  *   extern void WINAPI wine_call_to_16_regs_short( const CONTEXT86 *context, int nb_args );
 439  *   extern void WINAPI wine_call_to_16_regs_long ( const CONTEXT86 *context, int nb_args );
 440  *
 441  * These routines can be called directly from 32-bit code.
 442  *
 443  * All routines expect that the 16-bit stack contents (arguments) were
 444  * already set up by the caller; nb_args must contain the number of bytes
 445  * to be conserved.  The 16-bit SS:SP will be set accordinly.
 446  *
 447  * All other registers are either taken from the CONTEXT86 structure
 448  * or else set to default values.  The target routine address is either
 449  * given directly or taken from the CONTEXT86.
 450  *
 451  * If you want to call a 16-bit routine taking only standard argument types
 452  * (WORD and LONG), you can also have an appropriate argument conversion
 453  * stub automatically generated (see BuildCallTo16); you'd then call this
 454  * stub, which in turn would prepare the 16-bit stack and call the appropiate
 455  * core routine.
 456  *
 457  */
 458 static void BuildCallTo16Core( FILE *outfile, int short_ret, int reg_func )
 459 {
 460     char *name = reg_func == 2 ? "regs_long" :
 461                  reg_func == 1 ? "regs_short" :
 462                  short_ret? "word" : "long";
 463
 464     /* Function header */
 465     if (reg_func == 2) function_header( outfile, "wine_call_to_16_regs_long" );
 466     else if (reg_func == 1) function_header( outfile, "wine_call_to_16_regs_short" );
 467     else if (short_ret) function_header( outfile, "wine_call_to_16_word" );
 468     else function_header( outfile, "wine_call_to_16_long" );
 469
 470     /* Function entry sequence */
 471     fprintf( outfile, "\tpushl %%ebp\n" );
 472     fprintf( outfile, "\tmovl %%esp, %%ebp\n" );
 473
 474     /* Save the 32-bit registers */
 475     fprintf( outfile, "\tpushl %%ebx\n" );
 476     fprintf( outfile, "\tpushl %%ecx\n" );
 477     fprintf( outfile, "\tpushl %%edx\n" );
 478     fprintf( outfile, "\tpushl %%esi\n" );
 479     fprintf( outfile, "\tpushl %%edi\n" );
 480
 481     if ( UsePIC )
 482     {
 483         /* Get Global Offset Table into %ebx */
 484         fprintf( outfile, "\tcall .Lwine_call_to_16_%s.getgot1\n", name );
 485         fprintf( outfile, ".Lwine_call_to_16_%s.getgot1:\n", name );
 486         fprintf( outfile, "\tpopl %%ebx\n" );
 487         fprintf( outfile, "\taddl $_GLOBAL_OFFSET_TABLE_+[.-.Lwine_call_to_16_%s.getgot1], %%ebx\n", name );
 488     }
 489
 490     /* Enter Win16 Mutex */
 491     if ( UsePIC )
 492         fprintf( outfile, "\tcall " PREFIX "_EnterWin16Lock@PLT\n" );
 493     else
 494         fprintf( outfile, "\tcall " PREFIX "_EnterWin16Lock\n" );
 495
 496     /* Print debugging info */
 497     if (debugging)
 498     {
 499         /* Push flags, number of arguments, and target */
 500         fprintf( outfile, "\tpushl $%d\n", reg_func );
 501         fprintf( outfile, "\tpushl 12(%%ebp)\n" );
 502         fprintf( outfile, "\tpushl  8(%%ebp)\n" );
 503
 504         if ( UsePIC )
 505             fprintf( outfile, "\tcall " PREFIX "RELAY_DebugCallTo16@PLT\n" );
 506         else
 507             fprintf( outfile, "\tcall " PREFIX "RELAY_DebugCallTo16\n" );
 508
 509         fprintf( outfile, "\taddl $12, %%esp\n" );
 510     }
 511
 512     /* Get return address */
 513     if ( UsePIC )
 514     {
 515         fprintf( outfile, "\tmovl " PREFIX "CallTo16_RetAddr@GOT(%%ebx), %%ecx\n" );
 516         fprintf( outfile, "\tmovl " PREFIX "(%%ecx), %%ecx\n" );
 517     }
 518     else
 519         fprintf( outfile, "\tmovl " PREFIX "CallTo16_RetAddr, %%ecx\n" );
 520
 521     /* Call the actual CallTo16 routine (simulate a lcall) */
 522     fprintf( outfile, "\tpushl %%cs\n" );
 523     fprintf( outfile, "\tcall .Lwine_call_to_16_%s\n", name );
 524
 525     if ( !reg_func )
 526     {
 527         /* Convert and push return value */
 528         if ( short_ret )
 529         {
 530             fprintf( outfile, "\tmovzwl %%ax, %%eax\n" );
 531             fprintf( outfile, "\tpushl %%eax\n" );
 532         }
 533         else
 534         {
 535             fprintf( outfile, "\tshll $16,%%edx\n" );
 536             fprintf( outfile, "\tmovw %%ax,%%dx\n" );
 537             fprintf( outfile, "\tpushl %%edx\n" );
 538         }
 539     }
 540     else
 541     {
 542         /*
 543          * Modify CONTEXT86 structure to contain new values
 544          *
 545          * NOTE:  We restore only EAX, EBX, EDX, EDX, EBP, and ESP.
 546          *        The segment registers as well as ESI and EDI should
 547          *        not be modified by a well-behaved 16-bit routine in
 548          *        any case.  [If necessary, we could restore them as well,
 549          *        at the cost of a somewhat less efficient return path.]
 550          */
 551
 552         fprintf( outfile, "\tmovl %d(%%esp), %%edi\n", STACK32OFFSET(target)-12 );
 553         fprintf( outfile, "\tmovl %%eax, %d(%%edi)\n", CONTEXTOFFSET(Eax) );
 554         fprintf( outfile, "\tmovl %%ebx, %d(%%edi)\n", CONTEXTOFFSET(Ebx) );
 555         fprintf( outfile, "\tmovl %%ecx, %d(%%edi)\n", CONTEXTOFFSET(Ecx) );
 556         fprintf( outfile, "\tmovl %%edx, %d(%%edi)\n", CONTEXTOFFSET(Edx) );
 557         fprintf( outfile, "\tmovl %%ebp, %d(%%edi)\n", CONTEXTOFFSET(Ebp) );
 558         fprintf( outfile, "\tmovl %%esi, %d(%%edi)\n", CONTEXTOFFSET(Esp) );
 559                  /* The return glue code saved %esp into %esi */
 560
 561         fprintf( outfile, "\tpushl %%edi\n" );
 562     }
 563
 564     if ( UsePIC )
 565     {
 566         /* Get Global Offset Table into %ebx (might have been overwritten) */
 567         fprintf( outfile, "\tcall .Lwine_call_to_16_%s.getgot2\n", name );
 568         fprintf( outfile, ".Lwine_call_to_16_%s.getgot2:\n", name );
 569         fprintf( outfile, "\tpopl %%ebx\n" );
 570         fprintf( outfile, "\taddl $_GLOBAL_OFFSET_TABLE_+[.-.Lwine_call_to_16_%s.getgot2], %%ebx\n", name );
 571     }
 572
 573     /* Print debugging info */
 574     if (debugging)
 575     {
 576         fprintf( outfile, "\tpushl $%d\n", reg_func );
 577
 578         if ( UsePIC )
 579             fprintf( outfile, "\tcall " PREFIX "RELAY_DebugCallTo16Ret@PLT\n" );
 580         else
 581             fprintf( outfile, "\tcall " PREFIX "RELAY_DebugCallTo16Ret\n" );
 582
 583         fprintf( outfile, "\taddl $4, %%esp\n" );
 584     }
 585
 586     /* Leave Win16 Mutex */
 587     if ( UsePIC )
 588         fprintf( outfile, "\tcall " PREFIX "_LeaveWin16Lock@PLT\n" );
 589     else
 590         fprintf( outfile, "\tcall " PREFIX "_LeaveWin16Lock\n" );
 591
 592     /* Get return value */
 593     fprintf( outfile, "\tpopl %%eax\n" );
 594
 595     /* Restore the 32-bit registers */
 596     fprintf( outfile, "\tpopl %%edi\n" );
 597     fprintf( outfile, "\tpopl %%esi\n" );
 598     fprintf( outfile, "\tpopl %%edx\n" );
 599     fprintf( outfile, "\tpopl %%ecx\n" );
 600     fprintf( outfile, "\tpopl %%ebx\n" );
 601
 602     /* Function exit sequence */
 603     fprintf( outfile, "\tpopl %%ebp\n" );
 604     fprintf( outfile, "\tret $8\n" );
 605
 606
 607     /* Start of the actual CallTo16 routine */
 608
 609     fprintf( outfile, ".Lwine_call_to_16_%s:\n", name );
 610
 611     /* Complete STACK32FRAME */
 612     fprintf( outfile, "\t.byte 0x64\n\tpushl (%d)\n", STACKOFFSET );
 613     fprintf( outfile, "\tmovl %%esp,%%edx\n" );
 614
 615     /* Switch to the 16-bit stack */
 616 #ifdef __svr4__
 617     fprintf( outfile,"\tdata16\n");
 618 #endif
 619     fprintf( outfile, "\t.byte 0x64\n\tmovw (%d),%%ss\n", STACKOFFSET + 2);
 620     fprintf( outfile, "\t.byte 0x64\n\tmovw (%d),%%sp\n", STACKOFFSET );
 621     fprintf( outfile, "\t.byte 0x64\n\tmovl %%edx,(%d)\n", STACKOFFSET );
 622
 623     /* Make %bp point to the previous stackframe (built by CallFrom16) */
 624     fprintf( outfile, "\tmovzwl %%sp,%%ebp\n" );
 625     fprintf( outfile, "\tleal %d(%%ebp),%%ebp\n", STACK16OFFSET(bp) );
 626
 627     /* Add the specified offset to the new sp */
 628     fprintf( outfile, "\tsubw %d(%%edx), %%sp\n", STACK32OFFSET(nb_args) );
 629
 630     /* Push the return address
 631      * With sreg suffix, we push 16:16 address (normal lret)
 632      * With lreg suffix, we push 16:32 address (0x66 lret, for KERNEL32_45)
 633      */
 634     if (reg_func != 2)
 635         fprintf( outfile, "\tpushl %%ecx\n" );
 636     else
 637     {
 638         fprintf( outfile, "\tshldl $16, %%ecx, %%eax\n" );
 639         fprintf( outfile, "\tpushw $0\n" );
 640         fprintf( outfile, "\tpushw %%ax\n" );
 641         fprintf( outfile, "\tpushw $0\n" );
 642         fprintf( outfile, "\tpushw %%cx\n" );
 643     }
 644
 645     if (reg_func)
 646     {
 647         /* Push the called routine address */
 648         fprintf( outfile, "\tmovl %d(%%edx),%%edx\n", STACK32OFFSET(target) );
 649         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(SegCs) );
 650         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(Eip) );
 651
 652         /* Get the registers */
 653         fprintf( outfile, "\tpushw %d(%%edx)\n", CONTEXTOFFSET(SegDs) );
 654         fprintf( outfile, "\tmovl %d(%%edx),%%eax\n", CONTEXTOFFSET(SegEs) );
 655         fprintf( outfile, "\tmovw %%ax,%%es\n" );
 656         fprintf( outfile, "\tmovl %d(%%edx),%%eax\n", CONTEXTOFFSET(SegFs) );
 657         fprintf( outfile, "\tmovw %%ax,%%fs\n" );
 658         fprintf( outfile, "\tmovl %d(%%edx),%%ebp\n", CONTEXTOFFSET(Ebp) );
 659         fprintf( outfile, "\tmovl %d(%%edx),%%esi\n", CONTEXTOFFSET(Esi) );
 660         fprintf( outfile, "\tmovl %d(%%edx),%%edi\n", CONTEXTOFFSET(Edi) );
 661         fprintf( outfile, "\tmovl %d(%%edx),%%eax\n", CONTEXTOFFSET(Eax) );
 662         fprintf( outfile, "\tmovl %d(%%edx),%%ebx\n", CONTEXTOFFSET(Ebx) );
 663         fprintf( outfile, "\tmovl %d(%%edx),%%ecx\n", CONTEXTOFFSET(Ecx) );
 664         fprintf( outfile, "\tmovl %d(%%edx),%%edx\n", CONTEXTOFFSET(Edx) );
 665
 666         /* Get the 16-bit ds */
 667         fprintf( outfile, "\tpopw %%ds\n" );
 668     }
 669     else  /* not a register function */
 670     {
 671         /* Push the called routine address */
 672         fprintf( outfile, "\tpushl %d(%%edx)\n", STACK32OFFSET(target) );
 673
 674         /* Set %fs to the value saved by the last CallFrom16 */
 675         fprintf( outfile, "\tmovw %d(%%ebp),%%ax\n", STACK16OFFSET(fs)-STACK16OFFSET(bp) );
 676         fprintf( outfile, "\tmovw %%ax,%%fs\n" );
 677
 678         /* Set %ds and %es (and %ax just in case) equal to %ss */
 679         fprintf( outfile, "\tmovw %%ss,%%ax\n" );
 680         fprintf( outfile, "\tmovw %%ax,%%ds\n" );
 681         fprintf( outfile, "\tmovw %%ax,%%es\n" );
 682     }
 683
 684     /* Jump to the called routine */
 685     fprintf( outfile, "\t.byte 0x66\n" );
 686     fprintf( outfile, "\tlret\n" );
 687 }
 688
 689
 690 /*******************************************************************
 691  *         BuildRet16Func
 692  *
 693  * Build the return code for 16-bit callbacks
 694  */
 695 static void BuildRet16Func( FILE *outfile )
 696 {
 697     /*
 698      *  Note: This must reside in the .data section to allow
 699      *        run-time relocation of the SYSLEVEL_Win16CurrentTeb symbol
 700      */
 701
 702     function_header( outfile, "CallTo16_Ret" );
 703
 704     /* Save %esp into %esi */
 705     fprintf( outfile, "\tmovl %%esp,%%esi\n" );
 706
 707     /* Restore 32-bit segment registers */
 708
 709     fprintf( outfile, "\t.byte 0x2e\n\tmovl " PREFIX "CallTo16_DataSelector-" PREFIX "Call16_Ret_Start,%%edi\n" );
 710 #ifdef __svr4__
 711     fprintf( outfile, "\tdata16\n");
 712 #endif
 713     fprintf( outfile, "\tmovw %%di,%%ds\n" );
 714 #ifdef __svr4__
 715     fprintf( outfile, "\tdata16\n");
 716 #endif
 717     fprintf( outfile, "\tmovw %%di,%%es\n" );
 718
 719     fprintf( outfile, "\tmovw " PREFIX "SYSLEVEL_Win16CurrentTeb,%%fs\n" );
 720
 721     /* Restore the 32-bit stack */
 722
 723 #ifdef __svr4__
 724     fprintf( outfile, "\tdata16\n");
 725 #endif
 726     fprintf( outfile, "\tmovw %%di,%%ss\n" );
 727     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d),%%esp\n", STACKOFFSET );
 728     fprintf( outfile, "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 729
 730     /* Return to caller */
 731
 732     fprintf( outfile, "\tlret\n" );
 733
 734     /* Declare the return address and data selector variables */
 735
 736     fprintf( outfile, "\n\t.align %d\n", get_alignment(4) );
 737     fprintf( outfile, "\t.globl " PREFIX "CallTo16_DataSelector\n" );
 738     fprintf( outfile, PREFIX "CallTo16_DataSelector:\t.long 0\n" );
 739     fprintf( outfile, "\t.globl " PREFIX "CallTo16_RetAddr\n" );
 740     fprintf( outfile, PREFIX "CallTo16_RetAddr:\t.long 0\n" );
 741 }
 742
 743
 744 /*******************************************************************
 745  *         BuildCallTo32CBClient
 746  *
 747  * Call a CBClient relay stub from 32-bit code (KERNEL.620).
 748  *
 749  * Since the relay stub is itself 32-bit, this should not be a problem;
 750  * unfortunately, the relay stubs are expected to switch back to a
 751  * 16-bit stack (and 16-bit code) after completion :-(
 752  *
 753  * This would conflict with our 16- vs. 32-bit stack handling, so
 754  * we simply switch *back* to our 32-bit stack before returning to
 755  * the caller ...
 756  *
 757  * The CBClient relay stub expects to be called with the following
 758  * 16-bit stack layout, and with ebp and ebx pointing into the 16-bit
 759  * stack at the designated places:
 760  *
 761  *    ...
 762  *  (ebp+14) original arguments to the callback routine
 763  *  (ebp+10) far return address to original caller
 764  *  (ebp+6)  Thunklet target address
 765  *  (ebp+2)  Thunklet relay ID code
 766  *  (ebp)    BP (saved by CBClientGlueSL)
 767  *  (ebp-2)  SI (saved by CBClientGlueSL)
 768  *  (ebp-4)  DI (saved by CBClientGlueSL)
 769  *  (ebp-6)  DS (saved by CBClientGlueSL)
 770  *
 771  *   ...     buffer space used by the 16-bit side glue for temp copies
 772  *
 773  *  (ebx+4)  far return address to 16-bit side glue code
 774  *  (ebx)    saved 16-bit ss:sp (pointing to ebx+4)
 775  *
 776  * The 32-bit side glue code accesses both the original arguments (via ebp)
 777  * and the temporary copies prepared by the 16-bit side glue (via ebx).
 778  * After completion, the stub will load ss:sp from the buffer at ebx
 779  * and perform a far return to 16-bit code.
 780  *
 781  * To trick the relay stub into returning to us, we replace the 16-bit
 782  * return address to the glue code by a cs:ip pair pointing to our
 783  * return entry point (the original return address is saved first).
 784  * Our return stub thus called will then reload the 32-bit ss:esp and
 785  * return to 32-bit code (by using and ss:esp value that we have also
 786  * pushed onto the 16-bit stack before and a cs:eip values found at
 787  * that position on the 32-bit stack).  The ss:esp to be restored is
 788  * found relative to the 16-bit stack pointer at:
 789  *
 790  *  (ebx-4)   ss  (flat)
 791  *  (ebx-8)   sp  (32-bit stack pointer)
 792  *
 793  * The second variant of this routine, CALL32_CBClientEx, which is used
 794  * to implement KERNEL.621, has to cope with yet another problem: Here,
 795  * the 32-bit side directly returns to the caller of the CBClient thunklet,
 796  * restoring registers saved by CBClientGlueSL and cleaning up the stack.
 797  * As we have to return to our 32-bit code first, we have to adapt the
 798  * layout of our temporary area so as to include values for the registers
 799  * that are to be restored, and later (in the implementation of KERNEL.621)
 800  * we *really* restore them. The return stub restores DS, DI, SI, and BP
 801  * from the stack, skips the next 8 bytes (CBClient relay code / target),
 802  * and then performs a lret NN, where NN is the number of arguments to be
 803  * removed. Thus, we prepare our temporary area as follows:
 804  *
 805  *     (ebx+22) 16-bit cs  (this segment)
 806  *     (ebx+20) 16-bit ip  ('16-bit' return entry point)
 807  *     (ebx+16) 32-bit ss  (flat)
 808  *     (ebx+12) 32-bit sp  (32-bit stack pointer)
 809  *     (ebx+10) 16-bit bp  (points to ebx+24)
 810  *     (ebx+8)  16-bit si  (ignored)
 811  *     (ebx+6)  16-bit di  (ignored)
 812  *     (ebx+4)  16-bit ds  (we actually use the flat DS here)
 813  *     (ebx+2)  16-bit ss  (16-bit stack segment)
 814  *     (ebx+0)  16-bit sp  (points to ebx+4)
 815  *
 816  * Note that we ensure that DS is not changed and remains the flat segment,
 817  * and the 32-bit stack pointer our own return stub needs fits just
 818  * perfectly into the 8 bytes that are skipped by the Windows stub.
 819  * One problem is that we have to determine the number of removed arguments,
 820  * as these have to be really removed in KERNEL.621. Thus, the BP value
 821  * that we place in the temporary area to be restored, contains the value
 822  * that SP would have if no arguments were removed. By comparing the actual
 823  * value of SP with this value in our return stub we can compute the number
 824  * of removed arguments. This is then returned to KERNEL.621.
 825  *
 826  * The stack layout of this function:
 827  * (ebp+20)  nArgs     pointer to variable receiving nr. of args (Ex only)
 828  * (ebp+16)  esi       pointer to caller's esi value
 829  * (ebp+12)  arg       ebp value to be set for relay stub
 830  * (ebp+8)   func      CBClient relay stub address
 831  * (ebp+4)   ret addr
 832  * (ebp)     ebp
 833  */
 834 static void BuildCallTo32CBClient( FILE *outfile, BOOL isEx )
 835 {
 836     char *name = isEx? "CBClientEx" : "CBClient";
 837     int size = isEx? 24 : 12;
 838
 839     /* Function header */
 840
 841     fprintf( outfile, "\n\t.align %d\n", get_alignment(4) );
 842 #ifdef USE_STABS
 843     fprintf( outfile, ".stabs \"CALL32_%s:F1\",36,0,0," PREFIX "CALL32_%s\n",
 844                       name, name );
 845 #endif
 846     fprintf( outfile, "\t.globl " PREFIX "CALL32_%s\n", name );
 847     fprintf( outfile, PREFIX "CALL32_%s:\n", name );
 848
 849     /* Entry code */
 850
 851     fprintf( outfile, "\tpushl %%ebp\n" );
 852     fprintf( outfile, "\tmovl %%esp,%%ebp\n" );
 853     fprintf( outfile, "\tpushl %%edi\n" );
 854     fprintf( outfile, "\tpushl %%esi\n" );
 855     fprintf( outfile, "\tpushl %%ebx\n" );
 856
 857     /* Get the 16-bit stack */
 858
 859     fprintf( outfile, "\t.byte 0x64\n\tmovl (%d),%%ebx\n", STACKOFFSET);
 860
 861     /* Convert it to a flat address */
 862
 863     fprintf( outfile, "\tshldl $16,%%ebx,%%eax\n" );
 864     fprintf( outfile, "\tandl $0xfff8,%%eax\n" );
 865     fprintf( outfile, "\tshrl $1,%%eax\n" );
 866     fprintf( outfile, "\tmovl " PREFIX "wine_ldt_copy(%%eax),%%esi\n" );
 867     fprintf( outfile, "\tmovw %%bx,%%ax\n" );
 868     fprintf( outfile, "\taddl %%eax,%%esi\n" );
 869
 870     /* Allocate temporary area (simulate STACK16_PUSH) */
 871
 872     fprintf( outfile, "\tpushf\n" );
 873     fprintf( outfile, "\tcld\n" );
 874     fprintf( outfile, "\tleal -%d(%%esi), %%edi\n", size );
 875     fprintf( outfile, "\tmovl $%d, %%ecx\n", sizeof(STACK16FRAME) );
 876     fprintf( outfile, "\trep\n\tmovsb\n" );
 877     fprintf( outfile, "\tpopf\n" );
 878
 879     fprintf( outfile, "\t.byte 0x64\n\tsubw $%d,(%d)\n", size, STACKOFFSET );
 880
 881     fprintf( outfile, "\tpushl %%edi\n" );  /* remember address */
 882
 883     /* Set up temporary area */
 884
 885     if ( !isEx )
 886     {
 887         fprintf( outfile, "\tleal 4(%%edi), %%edi\n" );
 888
 889         fprintf( outfile, "\tleal -8(%%esp), %%eax\n" );
 890         fprintf( outfile, "\tmovl %%eax, -8(%%edi)\n" );    /* 32-bit sp */
 891
 892         fprintf( outfile, "\tmovw %%ss, %%ax\n" );
 893         fprintf( outfile, "\tandl $0x0000ffff, %%eax\n" );
 894         fprintf( outfile, "\tmovl %%eax, -4(%%edi)\n" );    /* 32-bit ss */
 895
 896         fprintf( outfile, "\taddl $%d, %%ebx\n", sizeof(STACK16FRAME)-size+4 + 4 );
 897         fprintf( outfile, "\tmovl %%ebx, 0(%%edi)\n" );    /* 16-bit ss:sp */
 898
 899         fprintf( outfile, "\tmovl " PREFIX "CALL32_%s_RetAddr, %%eax\n", name );
 900         fprintf( outfile, "\tmovl %%eax, 4(%%edi)\n" );   /* overwrite return address */
 901     }
 902     else
 903     {
 904         fprintf( outfile, "\taddl $%d, %%ebx\n", sizeof(STACK16FRAME)-size+4 );
 905         fprintf( outfile, "\tmovl %%ebx, 0(%%edi)\n" );
 906
 907         fprintf( outfile, "\tmovw %%ds, %%ax\n" );
 908         fprintf( outfile, "\tmovw %%ax, 4(%%edi)\n" );
 909
 910         fprintf( outfile, "\taddl $20, %%ebx\n" );
 911         fprintf( outfile, "\tmovw %%bx, 10(%%edi)\n" );
 912
 913         fprintf( outfile, "\tleal -8(%%esp), %%eax\n" );
 914         fprintf( outfile, "\tmovl %%eax, 12(%%edi)\n" );
 915
 916         fprintf( outfile, "\tmovw %%ss, %%ax\n" );
 917         fprintf( outfile, "\tandl $0x0000ffff, %%eax\n" );
 918         fprintf( outfile, "\tmovl %%eax, 16(%%edi)\n" );
 919
 920         fprintf( outfile, "\tmovl " PREFIX "CALL32_%s_RetAddr, %%eax\n", name );
 921         fprintf( outfile, "\tmovl %%eax, 20(%%edi)\n" );
 922     }
 923
 924     /* Set up registers and call CBClient relay stub (simulating a far call) */
 925
 926     fprintf( outfile, "\tmovl 16(%%ebp), %%esi\n" );
 927     fprintf( outfile, "\tmovl (%%esi), %%esi\n" );
 928
 929     fprintf( outfile, "\tmovl %%edi, %%ebx\n" );
 930     fprintf( outfile, "\tmovl 8(%%ebp), %%eax\n" );
 931     fprintf( outfile, "\tmovl 12(%%ebp), %%ebp\n" );
 932
 933     fprintf( outfile, "\tpushl %%cs\n" );
 934     fprintf( outfile, "\tcall *%%eax\n" );
 935
 936     /* Return new esi value to caller */
 937
 938     fprintf( outfile, "\tmovl 32(%%esp), %%edi\n" );
 939     fprintf( outfile, "\tmovl %%esi, (%%edi)\n" );
 940
 941     /* Cleanup temporary area (simulate STACK16_POP) */
 942
 943     fprintf( outfile, "\tpop %%esi\n" );
 944
 945     fprintf( outfile, "\tpushf\n" );
 946     fprintf( outfile, "\tstd\n" );
 947     fprintf( outfile, "\tdec %%esi\n" );
 948     fprintf( outfile, "\tleal %d(%%esi), %%edi\n", size );
 949     fprintf( outfile, "\tmovl $%d, %%ecx\n", sizeof(STACK16FRAME) );
 950     fprintf( outfile, "\trep\n\tmovsb\n" );
 951     fprintf( outfile, "\tpopf\n" );
 952
 953     fprintf( outfile, "\t.byte 0x64\n\taddw $%d,(%d)\n", size, STACKOFFSET );
 954
 955     /* Return argument size to caller */
 956     if ( isEx )
 957     {
 958         fprintf( outfile, "\tmovl 32(%%esp), %%ebx\n" );
 959         fprintf( outfile, "\tmovl %%ebp, (%%ebx)\n" );
 960     }
 961
 962     /* Restore registers and return */
 963
 964     fprintf( outfile, "\tpopl %%ebx\n" );
 965     fprintf( outfile, "\tpopl %%esi\n" );
 966     fprintf( outfile, "\tpopl %%edi\n" );
 967     fprintf( outfile, "\tpopl %%ebp\n" );
 968     fprintf( outfile, "\tret\n" );
 969 }
 970
 971 static void BuildCallTo32CBClientRet( FILE *outfile, BOOL isEx )
 972 {
 973     char *name = isEx? "CBClientEx" : "CBClient";
 974
 975     /* '16-bit' return stub */
 976
 977     fprintf( outfile, "\n\t.globl " PREFIX "CALL32_%s_Ret\n", name );
 978     fprintf( outfile, PREFIX "CALL32_%s_Ret:\n", name );
 979
 980     if ( !isEx )
 981     {
 982         fprintf( outfile, "\tmovzwl %%sp, %%ebx\n" );
 983         fprintf( outfile, "\tlssl %%ss:-16(%%ebx), %%esp\n" );
 984     }
 985     else
 986     {
 987         fprintf( outfile, "\tmovzwl %%bp, %%ebx\n" );
 988         fprintf( outfile, "\tsubw %%bp, %%sp\n" );
 989         fprintf( outfile, "\tmovzwl %%sp, %%ebp\n" );
 990         fprintf( outfile, "\tlssl %%ss:-12(%%ebx), %%esp\n" );
 991     }
 992     fprintf( outfile, "\tlret\n" );
 993
 994     /* Declare the return address variable */
 995
 996     fprintf( outfile, "\n\t.globl " PREFIX "CALL32_%s_RetAddr\n", name );
 997     fprintf( outfile, PREFIX "CALL32_%s_RetAddr:\t.long 0\n", name );
 998 }
 999
1000
1001 /*******************************************************************
1002  *         BuildCallFrom32Regs
1003  *
1004  * Build a 32-bit-to-Wine call-back function for a 'register' function.
1005  * 'args' is the number of dword arguments.
1006  *
1007  * Stack layout:
1008  *   ...
1009  * (ebp+12)  first arg
1010  * (ebp+8)   ret addr to user code
1011  * (ebp+4)   ret addr to relay code
1012  * (ebp+0)   saved ebp
1013  * (ebp-128) buffer area to allow stack frame manipulation
1014  * (ebp-332) CONTEXT86 struct
1015  * (ebp-336) CONTEXT86 *argument
1016  *  ....     other arguments copied from (ebp+12)
1017  *
1018  * The entry point routine is called with a CONTEXT* extra argument,
1019  * following the normal args. In this context structure, EIP_reg
1020  * contains the return address to user code, and ESP_reg the stack
1021  * pointer on return (with the return address and arguments already
1022  * removed).
1023  */
1024 static void BuildCallFrom32Regs( FILE *outfile )
1025 {
1026     static const int STACK_SPACE = 128 + sizeof(CONTEXT86);
1027
1028     /* Function header */
1029
1030     function_header( outfile, "__wine_call_from_32_regs" );
1031
1032     /* Allocate some buffer space on the stack */
1033
1034     fprintf( outfile, "\tpushl %%ebp\n" );
1035     fprintf( outfile, "\tmovl %%esp,%%ebp\n ");
1036     fprintf( outfile, "\tleal -%d(%%esp), %%esp\n", STACK_SPACE );
1037
1038     /* Build the context structure */
1039
1040     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Eax) - STACK_SPACE );
1041     fprintf( outfile, "\tpushfl\n" );
1042     fprintf( outfile, "\tpopl %%eax\n" );
1043     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(EFlags) - STACK_SPACE );
1044     fprintf( outfile, "\tmovl 0(%%ebp),%%eax\n" );
1045     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Ebp) - STACK_SPACE );
1046     fprintf( outfile, "\tmovl %%ebx,%d(%%ebp)\n", CONTEXTOFFSET(Ebx) - STACK_SPACE );
1047     fprintf( outfile, "\tmovl %%ecx,%d(%%ebp)\n", CONTEXTOFFSET(Ecx) - STACK_SPACE );
1048     fprintf( outfile, "\tmovl %%edx,%d(%%ebp)\n", CONTEXTOFFSET(Edx) - STACK_SPACE );
1049     fprintf( outfile, "\tmovl %%esi,%d(%%ebp)\n", CONTEXTOFFSET(Esi) - STACK_SPACE );
1050     fprintf( outfile, "\tmovl %%edi,%d(%%ebp)\n", CONTEXTOFFSET(Edi) - STACK_SPACE );
1051
1052     fprintf( outfile, "\txorl %%eax,%%eax\n" );
1053     fprintf( outfile, "\tmovw %%cs,%%ax\n" );
1054     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegCs) - STACK_SPACE );
1055     fprintf( outfile, "\tmovw %%es,%%ax\n" );
1056     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegEs) - STACK_SPACE );
1057     fprintf( outfile, "\tmovw %%fs,%%ax\n" );
1058     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegFs) - STACK_SPACE );
1059     fprintf( outfile, "\tmovw %%gs,%%ax\n" );
1060     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegGs) - STACK_SPACE );
1061     fprintf( outfile, "\tmovw %%ss,%%ax\n" );
1062     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegSs) - STACK_SPACE );
1063     fprintf( outfile, "\tmovw %%ds,%%ax\n" );
1064     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(SegDs) - STACK_SPACE );
1065     fprintf( outfile, "\tmovw %%ax,%%es\n" );  /* set %es equal to %ds just in case */
1066
1067     fprintf( outfile, "\tmovl $0x%x,%%eax\n", CONTEXT86_FULL );
1068     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(ContextFlags) - STACK_SPACE );
1069
1070     fprintf( outfile, "\tmovl 8(%%ebp),%%eax\n" ); /* Get %eip at time of call */
1071     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Eip) - STACK_SPACE );
1072
1073     /* Transfer the arguments */
1074
1075     fprintf( outfile, "\tmovl 4(%%ebp),%%ebx\n" );   /* get relay code addr */
1076     fprintf( outfile, "\tpushl %%esp\n" );           /* push ptr to context struct */
1077     fprintf( outfile, "\tmovzbl 4(%%ebx),%%ecx\n" ); /* fetch number of args to copy */
1078     fprintf( outfile, "\tjecxz 1f\n" );
1079     fprintf( outfile, "\tsubl %%ecx,%%esp\n" );
1080     fprintf( outfile, "\tleal 12(%%ebp),%%esi\n" );  /* get %esp at time of call */
1081     fprintf( outfile, "\tmovl %%esp,%%edi\n" );
1082     fprintf( outfile, "\tshrl $2,%%ecx\n" );
1083     fprintf( outfile, "\tcld\n" );
1084     fprintf( outfile, "\trep\n\tmovsl\n" );  /* copy args */
1085
1086     fprintf( outfile, "1:\tmovzbl 5(%%ebx),%%eax\n" ); /* fetch number of args to remove */
1087     fprintf( outfile, "\tleal 12(%%ebp,%%eax),%%eax\n" );
1088     fprintf( outfile, "\tmovl %%eax,%d(%%ebp)\n", CONTEXTOFFSET(Esp) - STACK_SPACE );
1089
1090     /* Call the entry point */
1091
1092     fprintf( outfile, "\tcall *0(%%ebx)\n" );
1093
1094     /* Store %eip and %ebp onto the new stack */
1095
1096     fprintf( outfile, "\tmovl %d(%%ebp),%%edx\n", CONTEXTOFFSET(Esp) - STACK_SPACE );
1097     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(Eip) - STACK_SPACE );
1098     fprintf( outfile, "\tmovl %%eax,-4(%%edx)\n" );
1099     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(Ebp) - STACK_SPACE );
1100     fprintf( outfile, "\tmovl %%eax,-8(%%edx)\n" );
1101
1102     /* Restore the context structure */
1103
1104     /* Note: we don't bother to restore %cs, %ds and %ss
1105      *       changing them in 32-bit code is a recipe for disaster anyway
1106      */
1107     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(SegEs) - STACK_SPACE );
1108     fprintf( outfile, "\tmovw %%ax,%%es\n" );
1109     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(SegFs) - STACK_SPACE );
1110     fprintf( outfile, "\tmovw %%ax,%%fs\n" );
1111     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(SegGs) - STACK_SPACE );
1112     fprintf( outfile, "\tmovw %%ax,%%gs\n" );
1113
1114     fprintf( outfile, "\tmovl %d(%%ebp),%%edi\n", CONTEXTOFFSET(Edi) - STACK_SPACE );
1115     fprintf( outfile, "\tmovl %d(%%ebp),%%esi\n", CONTEXTOFFSET(Esi) - STACK_SPACE );
1116     fprintf( outfile, "\tmovl %d(%%ebp),%%edx\n", CONTEXTOFFSET(Edx) - STACK_SPACE );
1117     fprintf( outfile, "\tmovl %d(%%ebp),%%ecx\n", CONTEXTOFFSET(Ecx) - STACK_SPACE );
1118     fprintf( outfile, "\tmovl %d(%%ebp),%%ebx\n", CONTEXTOFFSET(Ebx) - STACK_SPACE );
1119
1120     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(EFlags) - STACK_SPACE );
1121     fprintf( outfile, "\tpushl %%eax\n" );
1122     fprintf( outfile, "\tpopfl\n" );
1123     fprintf( outfile, "\tmovl %d(%%ebp),%%eax\n", CONTEXTOFFSET(Eax) - STACK_SPACE );
1124
1125     fprintf( outfile, "\tmovl %d(%%ebp),%%ebp\n", CONTEXTOFFSET(Esp) - STACK_SPACE );
1126     fprintf( outfile, "\tleal -8(%%ebp),%%esp\n" );
1127     fprintf( outfile, "\tpopl %%ebp\n" );
1128     fprintf( outfile, "\tret\n" );
1129 }
1130
1131
1132 /*******************************************************************
1133  *         BuildRelays16
1134  *
1135  * Build all the 16-bit relay callbacks
1136  */
1137 void BuildRelays16( FILE *outfile )
1138 {
1139     char buffer[1024];
1140
1141     /* File header */
1142
1143     fprintf( outfile, "/* File generated automatically. Do not edit! */\n\n" );
1144     fprintf( outfile, "\t.text\n" );
1145
1146 #ifdef USE_STABS
1147     if (output_file_name)
1148     {
1149         getcwd(buffer, sizeof(buffer));
1150         fprintf( outfile, "\t.file\t\"%s\"\n", output_file_name );
1151
1152         /*
1153          * The stabs help the internal debugger as they are an indication that it
1154          * is sensible to step into a thunk/trampoline.
1155          */
1156         fprintf( outfile, ".stabs \"%s/\",100,0,0,Code_Start\n", buffer);
1157         fprintf( outfile, ".stabs \"%s\",100,0,0,Code_Start\n", output_file_name );
1158         fprintf( outfile, "Code_Start:\n\n" );
1159     }
1160 #endif
1161     fprintf( outfile, PREFIX"Call16_Start:\n" );
1162     fprintf( outfile, "\t.globl "PREFIX"Call16_Start\n" );
1163     fprintf( outfile, "\t.byte 0\n\n" );
1164
1165     /* Standard CallFrom16 routine (WORD return) */
1166     BuildCallFrom16Core( outfile, FALSE, FALSE, TRUE );
1167
1168     /* Standard CallFrom16 routine (DWORD return) */
1169     BuildCallFrom16Core( outfile, FALSE, FALSE, FALSE );
1170
1171     /* Register CallFrom16 routine */
1172     BuildCallFrom16Core( outfile, TRUE, FALSE, FALSE );
1173
1174     /* C16ThkSL CallFrom16 routine */
1175     BuildCallFrom16Core( outfile, FALSE, TRUE, FALSE );
1176
1177     /* Standard CallTo16 routine (WORD return) */
1178     BuildCallTo16Core( outfile, TRUE, FALSE );
1179
1180     /* Standard CallTo16 routine (DWORD return) */
1181     BuildCallTo16Core( outfile, FALSE, FALSE );
1182
1183     /* Register CallTo16 routine (16:16 retf) */
1184     BuildCallTo16Core( outfile, FALSE, 1 );
1185
1186     /* Register CallTo16 routine (16:32 retf) */
1187     BuildCallTo16Core( outfile, FALSE, 2 );
1188
1189     /* CBClientThunkSL routine */
1190     BuildCallTo32CBClient( outfile, FALSE );
1191
1192     /* CBClientThunkSLEx routine */
1193     BuildCallTo32CBClient( outfile, TRUE  );
1194
1195     fprintf( outfile, PREFIX"Call16_End:\n" );
1196     fprintf( outfile, "\t.globl "PREFIX"Call16_End\n" );
1197
1198 #ifdef USE_STABS
1199     fprintf( outfile, "\t.stabs \"\",100,0,0,.Letext\n");
1200     fprintf( outfile, ".Letext:\n");
1201
1202     /* Some versions of gdb need to have the filename data for
1203        each section, so output it again for the data section. */
1204     if (output_file_name)
1205     {
1206         fprintf( outfile, ".stabs \"%s/\",100,0,0,Data_Start\n", buffer);
1207         fprintf( outfile, ".stabs \"%s\",100,0,0,Data_Start\n", output_file_name );
1208         fprintf( outfile, "Data_Start:\n\n" );
1209     }
1210 #endif
1211
1212     /* The whole Call16_Ret segment must lie within the .data section */
1213     fprintf( outfile, "\n\t.data\n" );
1214     fprintf( outfile, "\t.globl " PREFIX "Call16_Ret_Start\n" );
1215     fprintf( outfile, PREFIX "Call16_Ret_Start:\n" );
1216
1217     /* Standard CallTo16 return stub */
1218     BuildRet16Func( outfile );
1219
1220     /* CBClientThunkSL return stub */
1221     BuildCallTo32CBClientRet( outfile, FALSE );
1222
1223     /* CBClientThunkSLEx return stub */
1224     BuildCallTo32CBClientRet( outfile, TRUE  );
1225
1226     /* End of Call16_Ret segment */
1227     fprintf( outfile, "\n\t.globl " PREFIX "Call16_Ret_End\n" );
1228     fprintf( outfile, PREFIX "Call16_Ret_End:\n" );
1229 }
1230
1231 /*******************************************************************
1232  *         BuildRelays32
1233  *
1234  * Build all the 32-bit relay callbacks
1235  */
1236 void BuildRelays32( FILE *outfile )
1237 {
1238     /* File header */
1239
1240     fprintf( outfile, "/* File generated automatically. Do not edit! */\n\n" );
1241     fprintf( outfile, "\t.text\n" );
1242
1243 #ifdef USE_STABS
1244     if (output_file_name)
1245     {
1246         char buffer[1024];
1247         getcwd(buffer, sizeof(buffer));
1248         fprintf( outfile, "\t.file\t\"%s\"\n", output_file_name );
1249
1250         /*
1251          * The stabs help the internal debugger as they are an indication that it
1252          * is sensible to step into a thunk/trampoline.
1253          */
1254         fprintf( outfile, ".stabs \"%s/\",100,0,0,Code_Start\n", buffer);
1255         fprintf( outfile, ".stabs \"%s\",100,0,0,Code_Start\n", output_file_name );
1256         fprintf( outfile, "Code_Start:\n\n" );
1257     }
1258 #endif
1259     /* 32-bit register entry point */
1260     BuildCallFrom32Regs( outfile );
1261
1262 #ifdef USE_STABS
1263     fprintf( outfile, "\t.stabs \"\",100,0,0,.Letext\n");
1264     fprintf( outfile, ".Letext:\n");
1265 #endif
1266 }
1267
1268 #else /* __i386__ */
1269
1270 void BuildRelays16( FILE *outfile )
1271 {
1272     fprintf( outfile, "/* File not used with this architecture. Do not edit! */\n\n" );
1273 }
1274
1275 void BuildRelays32( FILE *outfile )
1276 {
1277     fprintf( outfile, "/* File not used with this architecture. Do not edit! */\n\n" );
1278 }
1279
1280 #endif  /* __i386__ */
1281