tools/winebuild/relay.c

   1 /*
   2  * Relay calls helper routines
   3  *
   4  * Copyright 1993 Robert J. Amstadt
   5  * Copyright 1995 Martin von Loewis
   6  * Copyright 1995, 1996, 1997 Alexandre Julliard
   7  * Copyright 1997 Eric Youngdale
   8  * Copyright 1999 Ulrich Weigand
   9  *
  10  * This library is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2.1 of the License, or (at your option) any later version.
  14  *
  15  * This library is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * Lesser General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU Lesser General Public
  21  * License along with this library; if not, write to the Free Software
  22  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  23  */
  24
  25 #include "config.h"
  26 #include "wine/port.h"
  27
  28 #include <ctype.h>
  29 #include <stdarg.h>
  30
  31 #include "build.h"
  32
  33 /* offset of the stack pointer relative to %fs:(0) */
  34 #define STACKOFFSET 0xc0  /* FIELD_OFFSET(TEB,WOW32Reserved) */
  35
  36 /* fix this if the ntdll_thread_regs structure is changed */
  37 #define GS_OFFSET  0x1d8  /* FIELD_OFFSET(TEB,SystemReserved2) + FIELD_OFFSET(ntdll_thread_data,gs) */
  38
  39 #define DPMI_VIF_OFFSET      (0x1fc + 0) /* FIELD_OFFSET(TEB,GdiTebBatch) + FIELD_OFFSET(WINE_VM86_TEB_INFO,dpmi_vif) */
  40 #define VM86_PENDING_OFFSET  (0x1fc + 4) /* FIELD_OFFSET(TEB,GdiTebBatch) + FIELD_OFFSET(WINE_VM86_TEB_INFO,vm86_pending) */
  41
  42 static void function_header( const char *name )
  43 {
  44     output( "\n\t.align %d\n", get_alignment(4) );
  45     output( "\t%s\n", func_declaration(name) );
  46     output( "%s\n", asm_globl(name) );
  47 }
  48
  49
  50 /*******************************************************************
  51  *         BuildCallFrom16Core
  52  *
  53  * This routine builds the core routines used in 16->32 thunks:
  54  * CallFrom16Word, CallFrom16Long, CallFrom16Register, and CallFrom16Thunk.
  55  *
  56  * These routines are intended to be called via a far call (with 32-bit
  57  * operand size) from 16-bit code.  The 16-bit code stub must push %bp,
  58  * the 32-bit entry point to be called, and the argument conversion
  59  * routine to be used (see stack layout below).
  60  *
  61  * The core routine completes the STACK16FRAME on the 16-bit stack and
  62  * switches to the 32-bit stack.  Then, the argument conversion routine
  63  * is called; it gets passed the 32-bit entry point and a pointer to the
  64  * 16-bit arguments (on the 16-bit stack) as parameters. (You can either
  65  * use conversion routines automatically generated by BuildCallFrom16,
  66  * or write your own for special purposes.)
  67  *
  68  * The conversion routine must call the 32-bit entry point, passing it
  69  * the converted arguments, and return its return value to the core.
  70  * After the conversion routine has returned, the core switches back
  71  * to the 16-bit stack, converts the return value to the DX:AX format
  72  * (CallFrom16Long), and returns to the 16-bit call stub.  All parameters,
  73  * including %bp, are popped off the stack.
  74  *
  75  * The 16-bit call stub now returns to the caller, popping the 16-bit
  76  * arguments if necessary (pascal calling convention).
  77  *
  78  * In the case of a 'register' function, CallFrom16Register fills a
  79  * CONTEXT86 structure with the values all registers had at the point
  80  * the first instruction of the 16-bit call stub was about to be
  81  * executed.  A pointer to this CONTEXT86 is passed as third parameter
  82  * to the argument conversion routine, which typically passes it on
  83  * to the called 32-bit entry point.
  84  *
  85  * CallFrom16Thunk is a special variant used by the implementation of
  86  * the Win95 16->32 thunk functions C16ThkSL and C16ThkSL01 and is
  87  * implemented as follows:
  88  * On entry, the EBX register is set up to contain a flat pointer to the
  89  * 16-bit stack such that EBX+22 points to the first argument.
  90  * Then, the entry point is called, while EBP is set up to point
  91  * to the return address (on the 32-bit stack).
  92  * The called function returns with CX set to the number of bytes
  93  * to be popped of the caller's stack.
  94  *
  95  * Stack layout upon entry to the core routine (STACK16FRAME):
  96  *  ...           ...
  97  * (sp+24) word   first 16-bit arg
  98  * (sp+22) word   cs
  99  * (sp+20) word   ip
 100  * (sp+18) word   bp
 101  * (sp+14) long   32-bit entry point (reused for Win16 mutex recursion count)
 102  * (sp+12) word   ip of actual entry point (necessary for relay debugging)
 103  * (sp+8)  long   relay (argument conversion) function entry point
 104  * (sp+4)  long   cs of 16-bit entry point
 105  * (sp)    long   ip of 16-bit entry point
 106  *
 107  * Added on the stack:
 108  * (sp-2)  word   saved gs
 109  * (sp-4)  word   saved fs
 110  * (sp-6)  word   saved es
 111  * (sp-8)  word   saved ds
 112  * (sp-12) long   saved ebp
 113  * (sp-16) long   saved ecx
 114  * (sp-20) long   saved edx
 115  * (sp-24) long   saved previous stack
 116  */
 117 static void BuildCallFrom16Core( int reg_func, int thunk )
 118 {
 119     /* Function header */
 120     if (thunk) function_header( "__wine_call_from_16_thunk" );
 121     else if (reg_func) function_header( "__wine_call_from_16_regs" );
 122     else function_header( "__wine_call_from_16" );
 123
 124     /* Create STACK16FRAME (except STACK32FRAME link) */
 125     output( "\tpushw %%gs\n" );
 126     output( "\tpushw %%fs\n" );
 127     output( "\tpushw %%es\n" );
 128     output( "\tpushw %%ds\n" );
 129     output( "\tpushl %%ebp\n" );
 130     output( "\tpushl %%ecx\n" );
 131     output( "\tpushl %%edx\n" );
 132
 133     /* Save original EFlags register */
 134     if (reg_func) output( "\tpushfl\n" );
 135
 136     if ( UsePIC )
 137     {
 138         output( "\tcall 1f\n" );
 139         output( "1:\tpopl %%ecx\n" );
 140         output( "\t.byte 0x2e\n\tmovl %s-1b(%%ecx),%%edx\n", asm_name("CallTo16_DataSelector") );
 141     }
 142     else
 143         output( "\t.byte 0x2e\n\tmovl %s,%%edx\n", asm_name("CallTo16_DataSelector") );
 144
 145     /* Load 32-bit segment registers */
 146     output( "\tmovw %%dx, %%ds\n" );
 147     output( "\tmovw %%dx, %%es\n" );
 148
 149     if ( UsePIC )
 150         output( "\tmovw %s-1b(%%ecx), %%fs\n", asm_name("CallTo16_TebSelector") );
 151     else
 152         output( "\tmovw %s, %%fs\n", asm_name("CallTo16_TebSelector") );
 153
 154     output( "\t.byte 0x64\n\tmov (%d),%%gs\n", GS_OFFSET );
 155
 156     /* Translate STACK16FRAME base to flat offset in %edx */
 157     output( "\tmovw %%ss, %%dx\n" );
 158     output( "\tandl $0xfff8, %%edx\n" );
 159     output( "\tshrl $1, %%edx\n" );
 160     if (UsePIC)
 161     {
 162         output( "\taddl wine_ldt_copy_ptr-1b(%%ecx),%%edx\n" );
 163         output( "\tmovl (%%edx), %%edx\n" );
 164     }
 165     else
 166         output( "\tmovl %s(%%edx), %%edx\n", asm_name("wine_ldt_copy") );
 167     output( "\tmovzwl %%sp, %%ebp\n" );
 168     output( "\tleal %d(%%ebp,%%edx), %%edx\n", reg_func ? 0 : -4 );
 169
 170     /* Get saved flags into %ecx */
 171     if (reg_func) output( "\tpopl %%ecx\n" );
 172
 173     /* Get the 32-bit stack pointer from the TEB and complete STACK16FRAME */
 174     output( "\t.byte 0x64\n\tmovl (%d), %%ebp\n", STACKOFFSET );
 175     output( "\tpushl %%ebp\n" );
 176
 177     /* Switch stacks */
 178     output( "\t.byte 0x64\n\tmovw %%ss, (%d)\n", STACKOFFSET + 2 );
 179     output( "\t.byte 0x64\n\tmovw %%sp, (%d)\n", STACKOFFSET );
 180     output( "\tpushl %%ds\n" );
 181     output( "\tpopl %%ss\n" );
 182     output( "\tmovl %%ebp, %%esp\n" );
 183     output( "\taddl $0x20,%%ebp\n");  /* FIELD_OFFSET(STACK32FRAME,ebp) */
 184
 185
 186     /* At this point:
 187        STACK16FRAME is completely set up
 188        DS, ES, SS: flat data segment
 189        FS: current TEB
 190        ESP: points to last STACK32FRAME
 191        EBP: points to ebp member of last STACK32FRAME
 192        EDX: points to current STACK16FRAME
 193        ECX: contains saved flags
 194        all other registers: unchanged */
 195
 196     /* Special case: C16ThkSL stub */
 197     if ( thunk )
 198     {
 199         /* Set up registers as expected and call thunk */
 200         output( "\tleal 0x1a(%%edx),%%ebx\n" );  /* sizeof(STACK16FRAME)-22 */
 201         output( "\tleal -4(%%esp), %%ebp\n" );
 202
 203         output( "\tcall *0x26(%%edx)\n");  /* FIELD_OFFSET(STACK16FRAME,entry_point) */
 204
 205         /* Switch stack back */
 206         output( "\t.byte 0x64\n\tmovw (%d), %%ss\n", STACKOFFSET+2 );
 207         output( "\t.byte 0x64\n\tmovzwl (%d), %%esp\n", STACKOFFSET );
 208         output( "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 209
 210         /* Restore registers and return directly to caller */
 211         output( "\taddl $8, %%esp\n" );
 212         output( "\tpopl %%ebp\n" );
 213         output( "\tpopw %%ds\n" );
 214         output( "\tpopw %%es\n" );
 215         output( "\tpopw %%fs\n" );
 216         output( "\tpopw %%gs\n" );
 217         output( "\taddl $20, %%esp\n" );
 218
 219         output( "\txorb %%ch, %%ch\n" );
 220         output( "\tpopl %%ebx\n" );
 221         output( "\taddw %%cx, %%sp\n" );
 222         output( "\tpush %%ebx\n" );
 223
 224         output( "\t.byte 0x66\n" );
 225         output( "\tlret\n" );
 226
 227         return;
 228     }
 229
 230
 231     /* Build register CONTEXT */
 232     if ( reg_func )
 233     {
 234         output( "\tsubl $0x2cc,%%esp\n" );       /* sizeof(CONTEXT86) */
 235
 236         output( "\tmovl %%ecx,0xc0(%%esp)\n" );  /* EFlags */
 237
 238         output( "\tmovl %%eax,0xb0(%%esp)\n" );  /* Eax */
 239         output( "\tmovl %%ebx,0xa4(%%esp)\n" );  /* Ebx */
 240         output( "\tmovl %%esi,0xa0(%%esp)\n" );  /* Esi */
 241         output( "\tmovl %%edi,0x9c(%%esp)\n" );  /* Edi */
 242
 243         output( "\tmovl 0x0c(%%edx),%%eax\n");   /* FIELD_OFFSET(STACK16FRAME,ebp) */
 244         output( "\tmovl %%eax,0xb4(%%esp)\n" );  /* Ebp */
 245         output( "\tmovl 0x08(%%edx),%%eax\n");   /* FIELD_OFFSET(STACK16FRAME,ecx) */
 246         output( "\tmovl %%eax,0xac(%%esp)\n" );  /* Ecx */
 247         output( "\tmovl 0x04(%%edx),%%eax\n");   /* FIELD_OFFSET(STACK16FRAME,edx) */
 248         output( "\tmovl %%eax,0xa8(%%esp)\n" );  /* Edx */
 249
 250         output( "\tmovzwl 0x10(%%edx),%%eax\n"); /* FIELD_OFFSET(STACK16FRAME,ds) */
 251         output( "\tmovl %%eax,0x98(%%esp)\n" );  /* SegDs */
 252         output( "\tmovzwl 0x12(%%edx),%%eax\n"); /* FIELD_OFFSET(STACK16FRAME,es) */
 253         output( "\tmovl %%eax,0x94(%%esp)\n" );  /* SegEs */
 254         output( "\tmovzwl 0x14(%%edx),%%eax\n"); /* FIELD_OFFSET(STACK16FRAME,fs) */
 255         output( "\tmovl %%eax,0x90(%%esp)\n" );  /* SegFs */
 256         output( "\tmovzwl 0x16(%%edx),%%eax\n"); /* FIELD_OFFSET(STACK16FRAME,gs) */
 257         output( "\tmovl %%eax,0x8c(%%esp)\n" );  /* SegGs */
 258
 259         output( "\tmovzwl 0x2e(%%edx),%%eax\n"); /* FIELD_OFFSET(STACK16FRAME,cs) */
 260         output( "\tmovl %%eax,0xbc(%%esp)\n" );  /* SegCs */
 261         output( "\tmovzwl 0x2c(%%edx),%%eax\n"); /* FIELD_OFFSET(STACK16FRAME,ip) */
 262         output( "\tmovl %%eax,0xb8(%%esp)\n" );  /* Eip */
 263
 264         output( "\t.byte 0x64\n\tmovzwl (%d), %%eax\n", STACKOFFSET+2 );
 265         output( "\tmovl %%eax,0xc8(%%esp)\n" );  /* SegSs */
 266         output( "\t.byte 0x64\n\tmovzwl (%d), %%eax\n", STACKOFFSET );
 267         output( "\taddl $0x2c,%%eax\n");         /* FIELD_OFFSET(STACK16FRAME,ip) */
 268         output( "\tmovl %%eax,0xc4(%%esp)\n" );  /* Esp */
 269 #if 0
 270         output( "\tfsave 0x1c(%%esp)\n" ); /* FloatSave */
 271 #endif
 272
 273         /* Push address of CONTEXT86 structure -- popped by the relay routine */
 274         output( "\tmovl %%esp,%%eax\n" );
 275         output( "\tandl $~15,%%esp\n" );
 276         output( "\tsubl $4,%%esp\n" );
 277         output( "\tpushl %%eax\n" );
 278     }
 279     else
 280     {
 281         output( "\tsubl $8,%%esp\n" );
 282         output( "\tandl $~15,%%esp\n" );
 283         output( "\taddl $8,%%esp\n" );
 284     }
 285
 286     /* Call relay routine (which will call the API entry point) */
 287     output( "\tleal 0x30(%%edx),%%eax\n" ); /* sizeof(STACK16FRAME) */
 288     output( "\tpushl %%eax\n" );
 289     output( "\tpushl 0x26(%%edx)\n");  /* FIELD_OFFSET(STACK16FRAME,entry_point) */
 290     output( "\tcall *0x20(%%edx)\n");  /* FIELD_OFFSET(STACK16FRAME,relay) */
 291
 292     if ( reg_func )
 293     {
 294         output( "\tleal -748(%%ebp),%%ebx\n" ); /* sizeof(CONTEXT) + FIELD_OFFSET(STACK32FRAME,ebp) */
 295
 296         /* Switch stack back */
 297         output( "\t.byte 0x64\n\tmovw (%d), %%ss\n", STACKOFFSET+2 );
 298         output( "\t.byte 0x64\n\tmovzwl (%d), %%esp\n", STACKOFFSET );
 299         output( "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 300
 301         /* Get return address to CallFrom16 stub */
 302         output( "\taddw $0x14,%%sp\n" ); /* FIELD_OFFSET(STACK16FRAME,callfrom_ip)-4 */
 303         output( "\tpopl %%eax\n" );
 304         output( "\tpopl %%edx\n" );
 305
 306         /* Restore all registers from CONTEXT */
 307         output( "\tmovw 0xc8(%%ebx),%%ss\n");   /* SegSs */
 308         output( "\tmovl 0xc4(%%ebx),%%esp\n");  /* Esp */
 309         output( "\taddl $4, %%esp\n" );  /* room for final return address */
 310
 311         output( "\tpushw 0xbc(%%ebx)\n");  /* SegCs */
 312         output( "\tpushw 0xb8(%%ebx)\n");  /* Eip */
 313         output( "\tpushl %%edx\n" );
 314         output( "\tpushl %%eax\n" );
 315         output( "\tpushl 0xc0(%%ebx)\n");  /* EFlags */
 316         output( "\tpushl 0x98(%%ebx)\n");  /* SegDs */
 317
 318         output( "\tpushl 0x94(%%ebx)\n");  /* SegEs */
 319         output( "\tpopl %%es\n" );
 320         output( "\tpushl 0x90(%%ebx)\n");  /* SegFs */
 321         output( "\tpopl %%fs\n" );
 322         output( "\tpushl 0x8c(%%ebx)\n");  /* SegGs */
 323         output( "\tpopl %%gs\n" );
 324
 325         output( "\tmovl 0xb4(%%ebx),%%ebp\n");  /* Ebp */
 326         output( "\tmovl 0xa0(%%ebx),%%esi\n");  /* Esi */
 327         output( "\tmovl 0x9c(%%ebx),%%edi\n");  /* Edi */
 328         output( "\tmovl 0xb0(%%ebx),%%eax\n");  /* Eax */
 329         output( "\tmovl 0xa8(%%ebx),%%edx\n");  /* Edx */
 330         output( "\tmovl 0xac(%%ebx),%%ecx\n");  /* Ecx */
 331         output( "\tmovl 0xa4(%%ebx),%%ebx\n");  /* Ebx */
 332
 333         output( "\tpopl %%ds\n" );
 334         output( "\tpopfl\n" );
 335         output( "\tlret\n" );
 336     }
 337     else
 338     {
 339         /* Switch stack back */
 340         output( "\t.byte 0x64\n\tmovw (%d), %%ss\n", STACKOFFSET+2 );
 341         output( "\t.byte 0x64\n\tmovzwl (%d), %%esp\n", STACKOFFSET );
 342         output( "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 343
 344         /* Restore registers */
 345         output( "\tpopl %%edx\n" );
 346         output( "\tpopl %%ecx\n" );
 347         output( "\tpopl %%ebp\n" );
 348         output( "\tpopw %%ds\n" );
 349         output( "\tpopw %%es\n" );
 350         output( "\tpopw %%fs\n" );
 351         output( "\tpopw %%gs\n" );
 352
 353         /* Return to return stub which will return to caller */
 354         output( "\tlret $12\n" );
 355     }
 356     if (thunk) output_function_size( "__wine_call_from_16_thunk" );
 357     else if (reg_func) output_function_size( "__wine_call_from_16_regs" );
 358     else output_function_size( "__wine_call_from_16" );
 359 }
 360
 361
 362 /*******************************************************************
 363  *         BuildCallTo16Core
 364  *
 365  * This routine builds the core routines used in 32->16 thunks:
 366  *
 367  * extern DWORD WINAPI wine_call_to_16( FARPROC16 target, DWORD cbArgs, PEXCEPTION_HANDLER handler );
 368  * extern void WINAPI wine_call_to_16_regs( CONTEXT86 *context, DWORD cbArgs, PEXCEPTION_HANDLER handler );
 369  *
 370  * These routines can be called directly from 32-bit code.
 371  *
 372  * All routines expect that the 16-bit stack contents (arguments) and the
 373  * return address (segptr to CallTo16_Ret) were already set up by the
 374  * caller; nb_args must contain the number of bytes to be conserved.  The
 375  * 16-bit SS:SP will be set accordingly.
 376  *
 377  * All other registers are either taken from the CONTEXT86 structure
 378  * or else set to default values.  The target routine address is either
 379  * given directly or taken from the CONTEXT86.
 380  */
 381 static void BuildCallTo16Core( int reg_func )
 382 {
 383     const char *name = reg_func ? "wine_call_to_16_regs" : "wine_call_to_16";
 384
 385     /* Function header */
 386     function_header( name );
 387
 388     /* Function entry sequence */
 389     output_cfi( ".cfi_startproc" );
 390     output( "\tpushl %%ebp\n" );
 391     output_cfi( ".cfi_adjust_cfa_offset 4" );
 392     output_cfi( ".cfi_rel_offset %%ebp,0" );
 393     output( "\tmovl %%esp, %%ebp\n" );
 394     output_cfi( ".cfi_def_cfa_register %%ebp" );
 395
 396     /* Save the 32-bit registers */
 397     output( "\tpushl %%ebx\n" );
 398     output_cfi( ".cfi_rel_offset %%ebx,-4" );
 399     output( "\tpushl %%esi\n" );
 400     output_cfi( ".cfi_rel_offset %%esi,-8" );
 401     output( "\tpushl %%edi\n" );
 402     output_cfi( ".cfi_rel_offset %%edi,-12" );
 403     output( "\t.byte 0x64\n\tmov %%gs,(%d)\n", GS_OFFSET );
 404
 405     /* Setup exception frame */
 406     output( "\t.byte 0x64\n\tpushl (%d)\n", STACKOFFSET );
 407     output( "\tpushl 16(%%ebp)\n" ); /* handler */
 408     output( "\t.byte 0x64\n\tpushl (0)\n" );
 409     output( "\t.byte 0x64\n\tmovl %%esp,(0)\n" );
 410
 411     /* Call the actual CallTo16 routine (simulate a lcall) */
 412     output( "\tpushl %%cs\n" );
 413     output( "\tcall .L%s\n", name );
 414
 415     /* Remove exception frame */
 416     output( "\t.byte 0x64\n\tpopl (0)\n" );
 417     output( "\taddl $4, %%esp\n" );
 418     output( "\t.byte 0x64\n\tpopl (%d)\n", STACKOFFSET );
 419
 420     if ( !reg_func )
 421     {
 422         /* Convert return value */
 423         output( "\tandl $0xffff,%%eax\n" );
 424         output( "\tshll $16,%%edx\n" );
 425         output( "\torl %%edx,%%eax\n" );
 426     }
 427     else
 428     {
 429         /*
 430          * Modify CONTEXT86 structure to contain new values
 431          *
 432          * NOTE:  We restore only EAX, EBX, EDX, EDX, EBP, and ESP.
 433          *        The segment registers as well as ESI and EDI should
 434          *        not be modified by a well-behaved 16-bit routine in
 435          *        any case.  [If necessary, we could restore them as well,
 436          *        at the cost of a somewhat less efficient return path.]
 437          */
 438
 439         output( "\tmovl 0x14(%%esp),%%edi\n" ); /* FIELD_OFFSET(STACK32FRAME,target) - FIELD_OFFSET(STACK32FRAME,edi) */
 440                 /* everything above edi has been popped already */
 441
 442         output( "\tmovl %%eax,0xb0(%%edi)\n");  /* Eax */
 443         output( "\tmovl %%ebx,0xa4(%%edi)\n");  /* Ebx */
 444         output( "\tmovl %%ecx,0xac(%%edi)\n");  /* Ecx */
 445         output( "\tmovl %%edx,0xa8(%%edi)\n");  /* Edx */
 446         output( "\tmovl %%ebp,0xb4(%%edi)\n");  /* Ebp */
 447         output( "\tmovl %%esi,0xc4(%%edi)\n");  /* Esp */
 448                  /* The return glue code saved %esp into %esi */
 449     }
 450
 451     /* Restore the 32-bit registers */
 452     output( "\tpopl %%edi\n" );
 453     output_cfi( ".cfi_same_value %%edi" );
 454     output( "\tpopl %%esi\n" );
 455     output_cfi( ".cfi_same_value %%esi" );
 456     output( "\tpopl %%ebx\n" );
 457     output_cfi( ".cfi_same_value %%ebx" );
 458
 459     /* Function exit sequence */
 460     output( "\tpopl %%ebp\n" );
 461     output_cfi( ".cfi_def_cfa %%esp,4" );
 462     output_cfi( ".cfi_same_value %%ebp" );
 463     output( "\tret $12\n" );
 464     output_cfi( ".cfi_endproc" );
 465
 466
 467     /* Start of the actual CallTo16 routine */
 468
 469     output( ".L%s:\n", name );
 470
 471     /* Switch to the 16-bit stack */
 472     output( "\tmovl %%esp,%%edx\n" );
 473     output( "\t.byte 0x64\n\tmovw (%d),%%ss\n", STACKOFFSET + 2);
 474     output( "\t.byte 0x64\n\tmovw (%d),%%sp\n", STACKOFFSET );
 475     output( "\t.byte 0x64\n\tmovl %%edx,(%d)\n", STACKOFFSET );
 476
 477     /* Make %bp point to the previous stackframe (built by CallFrom16) */
 478     output( "\tmovzwl %%sp,%%ebp\n" );
 479     output( "\tleal 0x2a(%%ebp),%%ebp\n");  /* FIELD_OFFSET(STACK16FRAME,bp) */
 480
 481     /* Add the specified offset to the new sp */
 482     output( "\tsubw 0x2c(%%edx), %%sp\n");  /* FIELD_OFFSET(STACK32FRAME,nb_args) */
 483
 484     if (reg_func)
 485     {
 486         /* Push the called routine address */
 487         output( "\tmovl 0x28(%%edx),%%edx\n");  /* FIELD_OFFSET(STACK32FRAME,target) */
 488         output( "\tpushw 0xbc(%%edx)\n");  /* SegCs */
 489         output( "\tpushw 0xb8(%%edx)\n");  /* Eip */
 490
 491         /* Get the registers */
 492         output( "\tpushw 0x98(%%edx)\n");  /* SegDs */
 493         output( "\tpushl 0x94(%%edx)\n");  /* SegEs */
 494         output( "\tpopl %%es\n" );
 495         output( "\tpushl 0x90(%%edx)\n");  /* SegFs */
 496         output( "\tpopl %%fs\n" );
 497         output( "\tpushl 0x8c(%%edx)\n");  /* SegGs */
 498         output( "\tpopl %%gs\n" );
 499         output( "\tmovl 0xb4(%%edx),%%ebp\n");  /* Ebp */
 500         output( "\tmovl 0xa0(%%edx),%%esi\n");  /* Esi */
 501         output( "\tmovl 0x9c(%%edx),%%edi\n");  /* Edi */
 502         output( "\tmovl 0xb0(%%edx),%%eax\n");  /* Eax */
 503         output( "\tmovl 0xa4(%%edx),%%ebx\n");  /* Ebx */
 504         output( "\tmovl 0xac(%%edx),%%ecx\n");  /* Ecx */
 505         output( "\tmovl 0xa8(%%edx),%%edx\n");  /* Edx */
 506
 507         /* Get the 16-bit ds */
 508         output( "\tpopw %%ds\n" );
 509     }
 510     else  /* not a register function */
 511     {
 512         /* Push the called routine address */
 513         output( "\tpushl 0x28(%%edx)\n"); /* FIELD_OFFSET(STACK32FRAME,target) */
 514
 515         /* Set %fs and %gs to the value saved by the last CallFrom16 */
 516         output( "\tpushw -22(%%ebp)\n" ); /* FIELD_OFFSET(STACK16FRAME,fs)-FIELD_OFFSET(STACK16FRAME,bp) */
 517         output( "\tpopw %%fs\n" );
 518         output( "\tpushw -20(%%ebp)\n" ); /* FIELD_OFFSET(STACK16FRAME,gs)-FIELD_OFFSET(STACK16FRAME,bp) */
 519         output( "\tpopw %%gs\n" );
 520
 521         /* Set %ds and %es (and %ax just in case) equal to %ss */
 522         output( "\tmovw %%ss,%%ax\n" );
 523         output( "\tmovw %%ax,%%ds\n" );
 524         output( "\tmovw %%ax,%%es\n" );
 525     }
 526
 527     /* Jump to the called routine */
 528     output( "\t.byte 0x66\n" );
 529     output( "\tlret\n" );
 530
 531     /* Function footer */
 532     output_function_size( name );
 533 }
 534
 535
 536 /*******************************************************************
 537  *         BuildRet16Func
 538  *
 539  * Build the return code for 16-bit callbacks
 540  */
 541 static void BuildRet16Func(void)
 542 {
 543     function_header( "__wine_call_to_16_ret" );
 544
 545     /* Save %esp into %esi */
 546     output( "\tmovl %%esp,%%esi\n" );
 547
 548     /* Restore 32-bit segment registers */
 549
 550     output( "\t.byte 0x2e\n\tmovl %s", asm_name("CallTo16_DataSelector") );
 551     output( "-%s,%%edi\n", asm_name("__wine_call16_start") );
 552     output( "\tmovw %%di,%%ds\n" );
 553     output( "\tmovw %%di,%%es\n" );
 554
 555     output( "\t.byte 0x2e\n\tmov %s", asm_name("CallTo16_TebSelector") );
 556     output( "-%s,%%fs\n", asm_name("__wine_call16_start") );
 557
 558     output( "\t.byte 0x64\n\tmov (%d),%%gs\n", GS_OFFSET );
 559
 560     /* Restore the 32-bit stack */
 561
 562     output( "\tmovw %%di,%%ss\n" );
 563     output( "\t.byte 0x64\n\tmovl (%d),%%esp\n", STACKOFFSET );
 564
 565     /* Return to caller */
 566
 567     output( "\tlret\n" );
 568     output_function_size( "__wine_call_to_16_ret" );
 569 }
 570
 571
 572 /*******************************************************************
 573  *         BuildCallTo32CBClient
 574  *
 575  * Call a CBClient relay stub from 32-bit code (KERNEL.620).
 576  *
 577  * Since the relay stub is itself 32-bit, this should not be a problem;
 578  * unfortunately, the relay stubs are expected to switch back to a
 579  * 16-bit stack (and 16-bit code) after completion :-(
 580  *
 581  * This would conflict with our 16- vs. 32-bit stack handling, so
 582  * we simply switch *back* to our 32-bit stack before returning to
 583  * the caller ...
 584  *
 585  * The CBClient relay stub expects to be called with the following
 586  * 16-bit stack layout, and with ebp and ebx pointing into the 16-bit
 587  * stack at the designated places:
 588  *
 589  *    ...
 590  *  (ebp+14) original arguments to the callback routine
 591  *  (ebp+10) far return address to original caller
 592  *  (ebp+6)  Thunklet target address
 593  *  (ebp+2)  Thunklet relay ID code
 594  *  (ebp)    BP (saved by CBClientGlueSL)
 595  *  (ebp-2)  SI (saved by CBClientGlueSL)
 596  *  (ebp-4)  DI (saved by CBClientGlueSL)
 597  *  (ebp-6)  DS (saved by CBClientGlueSL)
 598  *
 599  *   ...     buffer space used by the 16-bit side glue for temp copies
 600  *
 601  *  (ebx+4)  far return address to 16-bit side glue code
 602  *  (ebx)    saved 16-bit ss:sp (pointing to ebx+4)
 603  *
 604  * The 32-bit side glue code accesses both the original arguments (via ebp)
 605  * and the temporary copies prepared by the 16-bit side glue (via ebx).
 606  * After completion, the stub will load ss:sp from the buffer at ebx
 607  * and perform a far return to 16-bit code.
 608  *
 609  * To trick the relay stub into returning to us, we replace the 16-bit
 610  * return address to the glue code by a cs:ip pair pointing to our
 611  * return entry point (the original return address is saved first).
 612  * Our return stub thus called will then reload the 32-bit ss:esp and
 613  * return to 32-bit code (by using and ss:esp value that we have also
 614  * pushed onto the 16-bit stack before and a cs:eip values found at
 615  * that position on the 32-bit stack).  The ss:esp to be restored is
 616  * found relative to the 16-bit stack pointer at:
 617  *
 618  *  (ebx-4)   ss  (flat)
 619  *  (ebx-8)   sp  (32-bit stack pointer)
 620  *
 621  * The second variant of this routine, CALL32_CBClientEx, which is used
 622  * to implement KERNEL.621, has to cope with yet another problem: Here,
 623  * the 32-bit side directly returns to the caller of the CBClient thunklet,
 624  * restoring registers saved by CBClientGlueSL and cleaning up the stack.
 625  * As we have to return to our 32-bit code first, we have to adapt the
 626  * layout of our temporary area so as to include values for the registers
 627  * that are to be restored, and later (in the implementation of KERNEL.621)
 628  * we *really* restore them. The return stub restores DS, DI, SI, and BP
 629  * from the stack, skips the next 8 bytes (CBClient relay code / target),
 630  * and then performs a lret NN, where NN is the number of arguments to be
 631  * removed. Thus, we prepare our temporary area as follows:
 632  *
 633  *     (ebx+22) 16-bit cs  (this segment)
 634  *     (ebx+20) 16-bit ip  ('16-bit' return entry point)
 635  *     (ebx+16) 32-bit ss  (flat)
 636  *     (ebx+12) 32-bit sp  (32-bit stack pointer)
 637  *     (ebx+10) 16-bit bp  (points to ebx+24)
 638  *     (ebx+8)  16-bit si  (ignored)
 639  *     (ebx+6)  16-bit di  (ignored)
 640  *     (ebx+4)  16-bit ds  (we actually use the flat DS here)
 641  *     (ebx+2)  16-bit ss  (16-bit stack segment)
 642  *     (ebx+0)  16-bit sp  (points to ebx+4)
 643  *
 644  * Note that we ensure that DS is not changed and remains the flat segment,
 645  * and the 32-bit stack pointer our own return stub needs fits just
 646  * perfectly into the 8 bytes that are skipped by the Windows stub.
 647  * One problem is that we have to determine the number of removed arguments,
 648  * as these have to be really removed in KERNEL.621. Thus, the BP value
 649  * that we place in the temporary area to be restored, contains the value
 650  * that SP would have if no arguments were removed. By comparing the actual
 651  * value of SP with this value in our return stub we can compute the number
 652  * of removed arguments. This is then returned to KERNEL.621.
 653  *
 654  * The stack layout of this function:
 655  * (ebp+20)  nArgs     pointer to variable receiving nr. of args (Ex only)
 656  * (ebp+16)  esi       pointer to caller's esi value
 657  * (ebp+12)  arg       ebp value to be set for relay stub
 658  * (ebp+8)   func      CBClient relay stub address
 659  * (ebp+4)   ret addr
 660  * (ebp)     ebp
 661  */
 662 static void BuildCallTo32CBClient( int isEx )
 663 {
 664     function_header( isEx ? "CALL32_CBClientEx" : "CALL32_CBClient" );
 665
 666     /* Entry code */
 667
 668     output_cfi( ".cfi_startproc" );
 669     output( "\tpushl %%ebp\n" );
 670     output_cfi( ".cfi_adjust_cfa_offset 4" );
 671     output_cfi( ".cfi_rel_offset %%ebp,0" );
 672     output( "\tmovl %%esp,%%ebp\n" );
 673     output_cfi( ".cfi_def_cfa_register %%ebp" );
 674     output( "\tpushl %%edi\n" );
 675     output_cfi( ".cfi_rel_offset %%edi,-4" );
 676     output( "\tpushl %%esi\n" );
 677     output_cfi( ".cfi_rel_offset %%esi,-8" );
 678     output( "\tpushl %%ebx\n" );
 679     output_cfi( ".cfi_rel_offset %%ebx,-12" );
 680
 681     /* Get pointer to temporary area and save the 32-bit stack pointer */
 682
 683     output( "\tmovl 16(%%ebp), %%ebx\n" );
 684     output( "\tleal -8(%%esp), %%eax\n" );
 685
 686     if ( !isEx )
 687         output( "\tmovl %%eax, -8(%%ebx)\n" );
 688     else
 689         output( "\tmovl %%eax, 12(%%ebx)\n" );
 690
 691     /* Set up registers and call CBClient relay stub (simulating a far call) */
 692
 693     output( "\tmovl 20(%%ebp), %%esi\n" );
 694     output( "\tmovl (%%esi), %%esi\n" );
 695
 696     output( "\tmovl 8(%%ebp), %%eax\n" );
 697     output( "\tmovl 12(%%ebp), %%ebp\n" );
 698
 699     output( "\tpushl %%cs\n" );
 700     output( "\tcall *%%eax\n" );
 701
 702     /* Return new esi value to caller */
 703
 704     output( "\tmovl 32(%%esp), %%edi\n" );
 705     output( "\tmovl %%esi, (%%edi)\n" );
 706
 707     /* Return argument size to caller */
 708     if ( isEx )
 709     {
 710         output( "\tmovl 36(%%esp), %%ebx\n" );
 711         output( "\tmovl %%ebp, (%%ebx)\n" );
 712     }
 713
 714     /* Restore registers and return */
 715
 716     output( "\tpopl %%ebx\n" );
 717     output_cfi( ".cfi_same_value %%ebx" );
 718     output( "\tpopl %%esi\n" );
 719     output_cfi( ".cfi_same_value %%esi" );
 720     output( "\tpopl %%edi\n" );
 721     output_cfi( ".cfi_same_value %%edi" );
 722     output( "\tpopl %%ebp\n" );
 723     output_cfi( ".cfi_def_cfa %%esp,4" );
 724     output_cfi( ".cfi_same_value %%ebp" );
 725     output( "\tret\n" );
 726     output_cfi( ".cfi_endproc" );
 727     output_function_size( isEx ? "CALL32_CBClientEx" : "CALL32_CBClient" );
 728
 729     /* '16-bit' return stub */
 730
 731     function_header( isEx ? "CALL32_CBClientEx_Ret" : "CALL32_CBClient_Ret" );
 732     if ( !isEx )
 733     {
 734         output( "\tmovzwl %%sp, %%ebx\n" );
 735         output( "\tlssl %%ss:-16(%%ebx), %%esp\n" );
 736     }
 737     else
 738     {
 739         output( "\tmovzwl %%bp, %%ebx\n" );
 740         output( "\tsubw %%bp, %%sp\n" );
 741         output( "\tmovzwl %%sp, %%ebp\n" );
 742         output( "\tlssl %%ss:-12(%%ebx), %%esp\n" );
 743     }
 744     output( "\tlret\n" );
 745     output_function_size( isEx ? "CALL32_CBClientEx_Ret" : "CALL32_CBClient_Ret" );
 746 }
 747
 748
 749 /*******************************************************************
 750  *         build_call_from_regs_x86
 751  *
 752  * Build a 32-bit-to-Wine call-back function for a 'register' function.
 753  * 'args' is the number of dword arguments.
 754  *
 755  * Stack layout:
 756  *   ...
 757  * (ebp+20)  first arg
 758  * (ebp+16)  ret addr to user code
 759  * (ebp+12)  func to call (relative to relay code ret addr)
 760  * (ebp+8)   number of args
 761  * (ebp+4)   ret addr to relay code
 762  * (ebp+0)   saved ebp
 763  * (ebp-128) buffer area to allow stack frame manipulation
 764  * (ebp-332) CONTEXT86 struct
 765  * (ebp-336) padding for stack alignment
 766  * (ebp-336-n) CONTEXT86 *argument
 767  *  ....     other arguments copied from (ebp+12)
 768  *
 769  * The entry point routine is called with a CONTEXT* extra argument,
 770  * following the normal args. In this context structure, EIP_reg
 771  * contains the return address to user code, and ESP_reg the stack
 772  * pointer on return (with the return address and arguments already
 773  * removed).
 774  */
 775 static void build_call_from_regs_x86(void)
 776 {
 777     static const int STACK_SPACE = 128 + 0x2cc /* sizeof(CONTEXT86) */;
 778
 779     /* Function header */
 780
 781     output( "\t.text\n" );
 782     function_header( "__wine_call_from_regs" );
 783
 784     /* Allocate some buffer space on the stack */
 785
 786     output_cfi( ".cfi_startproc" );
 787     output( "\tpushl %%ebp\n" );
 788     output_cfi( ".cfi_adjust_cfa_offset 4" );
 789     output_cfi( ".cfi_rel_offset %%ebp,0" );
 790     output( "\tmovl %%esp,%%ebp\n" );
 791     output_cfi( ".cfi_def_cfa_register %%ebp" );
 792     output( "\tleal -%d(%%esp),%%esp\n", STACK_SPACE );
 793
 794     /* Build the context structure */
 795
 796     output( "\tmovl %%eax,0xb0(%%esp)\n" );  /* Eax */
 797     output( "\tpushfl\n" );
 798     output( "\tpopl %%eax\n" );
 799     output( "\tmovl %%eax,0xc0(%%esp)\n");  /* EFlags */
 800     output( "\tmovl 0(%%ebp),%%eax\n" );
 801     output( "\tmovl %%eax,0xb4(%%esp)\n");  /* Ebp */
 802     output( "\tmovl %%ebx,0xa4(%%esp)\n");  /* Ebx */
 803     output( "\tmovl %%ecx,0xac(%%esp)\n");  /* Ecx */
 804     output( "\tmovl %%edx,0xa8(%%esp)\n");  /* Edx */
 805     output( "\tmovl %%esi,0xa0(%%esp)\n");  /* Esi */
 806     output( "\tmovl %%edi,0x9c(%%esp)\n");  /* Edi */
 807
 808     output( "\txorl %%eax,%%eax\n" );
 809     output( "\tmovw %%cs,%%ax\n" );
 810     output( "\tmovl %%eax,0xbc(%%esp)\n");  /* SegCs */
 811     output( "\tmovw %%es,%%ax\n" );
 812     output( "\tmovl %%eax,0x94(%%esp)\n");  /* SegEs */
 813     output( "\tmovw %%fs,%%ax\n" );
 814     output( "\tmovl %%eax,0x90(%%esp)\n");  /* SegFs */
 815     output( "\tmovw %%gs,%%ax\n" );
 816     output( "\tmovl %%eax,0x8c(%%esp)\n");  /* SegGs */
 817     output( "\tmovw %%ss,%%ax\n" );
 818     output( "\tmovl %%eax,0xc8(%%esp)\n");  /* SegSs */
 819     output( "\tmovw %%ds,%%ax\n" );
 820     output( "\tmovl %%eax,0x98(%%esp)\n");  /* SegDs */
 821     output( "\tmovw %%ax,%%es\n" );  /* set %es equal to %ds just in case */
 822
 823     output( "\tmovl $0x10007,0(%%esp)\n");  /* ContextFlags */
 824
 825     output( "\tmovl 16(%%ebp),%%eax\n" ); /* Get %eip at time of call */
 826     output( "\tmovl %%eax,0xb8(%%esp)\n");  /* Eip */
 827
 828     /* Transfer the arguments */
 829
 830     output( "\tmovl 8(%%ebp),%%ecx\n" );    /* fetch number of args to copy */
 831     output( "\tleal 4(,%%ecx,4),%%edx\n" ); /* add 4 for context arg */
 832     output( "\tsubl %%edx,%%esp\n" );
 833     output( "\tandl $~15,%%esp\n" );
 834     output( "\tleal 20(%%ebp),%%esi\n" );  /* get %esp at time of call */
 835     output( "\tmovl %%esp,%%edi\n" );
 836     output( "\ttest %%ecx,%%ecx\n" );
 837     output( "\tjz 1f\n" );
 838     output( "\tcld\n" );
 839     output( "\trep\n\tmovsl\n" );  /* copy args */
 840     output( "1:\tleal %d(%%ebp),%%eax\n", -STACK_SPACE );  /* get addr of context struct */
 841     output( "\tmovl %%eax,(%%edi)\n" );    /* and pass it as extra arg */
 842     output( "\tmovl %%esi,%d(%%ebp)\n", 0xc4 /* Esp */ - STACK_SPACE );
 843
 844     /* Call the entry point */
 845
 846     output( "\tmovl 4(%%ebp),%%eax\n" );   /* get relay code addr */
 847     output( "\taddl 12(%%ebp),%%eax\n" );
 848     output( "\tcall *%%eax\n" );
 849     output( "\tleal -%d(%%ebp),%%ecx\n", STACK_SPACE );
 850
 851     /* Restore the context structure */
 852
 853     output( "2:\tpushl 0x94(%%ecx)\n");     /* SegEs */
 854     output( "\tpopl %%es\n" );
 855     output( "\tpushl 0x90(%%ecx)\n");       /* SegFs */
 856     output( "\tpopl %%fs\n" );
 857     output( "\tpushl 0x8c(%%ecx)\n");       /* SegGs */
 858     output( "\tpopl %%gs\n" );
 859
 860     output( "\tmovl 0x9c(%%ecx),%%edi\n");  /* Edi */
 861     output( "\tmovl 0xa0(%%ecx),%%esi\n");  /* Esi */
 862     output( "\tmovl 0xa8(%%ecx),%%edx\n");  /* Edx */
 863     output( "\tmovl 0xa4(%%ecx),%%ebx\n");  /* Ebx */
 864     output( "\tmovl 0xb0(%%ecx),%%eax\n");  /* Eax */
 865     output( "\tmovl 0xb4(%%ecx),%%ebp\n");  /* Ebp */
 866
 867     output( "\tpushl 0xc8(%%ecx)\n");       /* SegSs */
 868     output( "\tpopl %%ss\n" );
 869     output( "\tmovl 0xc4(%%ecx),%%esp\n");  /* Esp */
 870
 871     output( "\tpushl 0xc0(%%ecx)\n");       /* EFlags */
 872     output( "\tpushl 0xbc(%%ecx)\n");       /* SegCs */
 873     output( "\tpushl 0xb8(%%ecx)\n");       /* Eip */
 874     output( "\tpushl 0x98(%%ecx)\n");       /* SegDs */
 875     output( "\tmovl 0xac(%%ecx),%%ecx\n");  /* Ecx */
 876
 877     output( "\tpopl %%ds\n" );
 878     output( "\tiret\n" );
 879     output_cfi( ".cfi_endproc" );
 880     output_function_size( "__wine_call_from_regs" );
 881
 882     function_header( "__wine_restore_regs" );
 883     output_cfi( ".cfi_startproc" );
 884     output( "\tmovl 4(%%esp),%%ecx\n" );
 885     output( "\tjmp 2b\n" );
 886     output_cfi( ".cfi_endproc" );
 887     output_function_size( "__wine_restore_regs" );
 888 }
 889
 890
 891 /*******************************************************************
 892  *         BuildPendingEventCheck
 893  *
 894  * Build a function that checks whether there are any
 895  * pending DPMI events.
 896  *
 897  * Stack layout:
 898  *
 899  * (sp+12) long   eflags
 900  * (sp+6)  long   cs
 901  * (sp+2)  long   ip
 902  * (sp)    word   fs
 903  *
 904  * On entry to function, fs register points to a valid TEB.
 905  * On exit from function, stack will be popped.
 906  */
 907 static void BuildPendingEventCheck(void)
 908 {
 909     /* Function header */
 910
 911     function_header( "DPMI_PendingEventCheck" );
 912
 913     /* Check for pending events. */
 914
 915     output( "\t.byte 0x64\n\ttestl $0xffffffff,(%d)\n", VM86_PENDING_OFFSET );
 916     output( "\tje %s\n", asm_name("DPMI_PendingEventCheck_Cleanup") );
 917     output( "\t.byte 0x64\n\ttestl $0xffffffff,(%d)\n", DPMI_VIF_OFFSET );
 918     output( "\tje %s\n", asm_name("DPMI_PendingEventCheck_Cleanup") );
 919
 920     /* Process pending events. */
 921
 922     output( "\tsti\n" );
 923
 924     /* Start cleanup. Restore fs register. */
 925
 926     output( "%s\n", asm_globl("DPMI_PendingEventCheck_Cleanup") );
 927     output( "\tpopw %%fs\n" );
 928
 929     /* Return from function. */
 930
 931     output( "%s\n", asm_globl("DPMI_PendingEventCheck_Return") );
 932     output( "\tiret\n" );
 933
 934     output_function_size( "DPMI_PendingEventCheck" );
 935 }
 936
 937
 938 /*******************************************************************
 939  *         output_asm_relays16
 940  *
 941  * Build all the 16-bit relay callbacks
 942  */
 943 void output_asm_relays16(void)
 944 {
 945     /* File header */
 946
 947     output( "\t.text\n" );
 948     output( "%s:\n\n", asm_name("__wine_spec_thunk_text_16") );
 949
 950     output( "%s\n", asm_globl("__wine_call16_start") );
 951
 952     /* Standard CallFrom16 routine */
 953     BuildCallFrom16Core( 0, 0 );
 954
 955     /* Register CallFrom16 routine */
 956     BuildCallFrom16Core( 1, 0 );
 957
 958     /* C16ThkSL CallFrom16 routine */
 959     BuildCallFrom16Core( 0, 1 );
 960
 961     /* Standard CallTo16 routine */
 962     BuildCallTo16Core( 0 );
 963
 964     /* Register CallTo16 routine */
 965     BuildCallTo16Core( 1 );
 966
 967     /* Standard CallTo16 return stub */
 968     BuildRet16Func();
 969
 970     /* CBClientThunkSL routine */
 971     BuildCallTo32CBClient( 0 );
 972
 973     /* CBClientThunkSLEx routine */
 974     BuildCallTo32CBClient( 1  );
 975
 976     /* Pending DPMI events check stub */
 977     BuildPendingEventCheck();
 978
 979     output( "%s\n", asm_globl("__wine_call16_end") );
 980     output_function_size( "__wine_spec_thunk_text_16" );
 981
 982     /* Declare the return address and data selector variables */
 983     output( "\n\t.data\n\t.align %d\n", get_alignment(4) );
 984     output( "%s\n\t.long 0\n", asm_globl("CallTo16_DataSelector") );
 985     output( "%s\n\t.long 0\n", asm_globl("CallTo16_TebSelector") );
 986
 987     output( "\t.text\n" );
 988     output( "%s:\n", asm_name("__wine_spec_thunk_text_32") );
 989     build_call_from_regs_x86();
 990     output_function_size( "__wine_spec_thunk_text_32" );
 991 }
 992
 993
 994 /*******************************************************************
 995  *         build_call_from_regs_x86_64
 996  *
 997  * Build the register saving code for a 'register' entry point.
 998  *
 999  * Stack layout:
1000  *   ...
1001  * (rsp+16)  first arg
1002  * (rsp+8)   ret addr to user code
1003  * (rsp)     ret addr to relay code
1004  * (rsp-128) buffer area to allow stack frame manipulation
1005  *
1006  * Parameters:
1007  *  %rcx     number of args
1008  *  %rdx     entry point
1009  */
1010 static void build_call_from_regs_x86_64(void)
1011 {
1012     static const int STACK_SPACE = 128 + 0x4d0; /* size of x86_64 context */
1013     int i;
1014
1015     /* Function header */
1016
1017     output( "\t.text\n" );
1018     function_header( "__wine_call_from_regs" );
1019
1020     output_cfi( ".cfi_startproc" );
1021     output( "\tsubq $%u,%%rsp\n", STACK_SPACE );
1022     output_cfi( ".cfi_adjust_cfa_offset %u", STACK_SPACE );
1023
1024     /* save registers into the context */
1025
1026     output( "\tmovq %%rax,0x78(%%rsp)\n" );
1027     output_cfi( ".cfi_rel_offset %%rax,0x78" );
1028     output( "\tmovq %u(%%rsp),%%rax\n", STACK_SPACE + 16 );  /* saved %rcx on stack */
1029     output( "\tmovq %%rax,0x80(%%rsp)\n" );
1030     output_cfi( ".cfi_rel_offset %%rcx,0x80" );
1031     output( "\tmovq %u(%%rsp),%%rax\n", STACK_SPACE + 24 );  /* saved %rdx on stack */
1032     output_cfi( ".cfi_rel_offset %%rdx,0x88" );
1033     output( "\tmovq %%rax,0x88(%%rsp)\n" );
1034     output( "\tmovq %%rbx,0x90(%%rsp)\n" );
1035     output_cfi( ".cfi_rel_offset %%rbx,0x90" );
1036     output( "\tleaq %u(%%rsp),%%rax\n", STACK_SPACE + 16 );
1037     output( "\tmovq %%rax,0x98(%%rsp)\n" );
1038     output( "\tmovq %%rbp,0xa0(%%rsp)\n" );
1039     output_cfi( ".cfi_rel_offset %%rbp,0xa0" );
1040     output( "\tmovq %%rsi,0xa8(%%rsp)\n" );
1041     output_cfi( ".cfi_rel_offset %%rsi,0xa8" );
1042     output( "\tmovq %%rdi,0xb0(%%rsp)\n" );
1043     output_cfi( ".cfi_rel_offset %%rdi,0xb0" );
1044     output( "\tmovq %%r8,0xb8(%%rsp)\n" );
1045     output_cfi( ".cfi_rel_offset %%r8,0xb8" );
1046     output( "\tmovq %%r9,0xc0(%%rsp)\n" );
1047     output_cfi( ".cfi_rel_offset %%r9,0xc0" );
1048     output( "\tmovq %%r10,0xc8(%%rsp)\n" );
1049     output_cfi( ".cfi_rel_offset %%r10,0xc8" );
1050     output( "\tmovq %%r11,0xd0(%%rsp)\n" );
1051     output_cfi( ".cfi_rel_offset %%r11,0xd0" );
1052     output( "\tmovq %%r12,0xd8(%%rsp)\n" );
1053     output_cfi( ".cfi_rel_offset %%r12,0xd8" );
1054     output( "\tmovq %%r13,0xe0(%%rsp)\n" );
1055     output_cfi( ".cfi_rel_offset %%r13,0xe0" );
1056     output( "\tmovq %%r14,0xe8(%%rsp)\n" );
1057     output_cfi( ".cfi_rel_offset %%r14,0xe8" );
1058     output( "\tmovq %%r15,0xf0(%%rsp)\n" );
1059     output_cfi( ".cfi_rel_offset %%r15,0xf0" );
1060     output( "\tmovq %u(%%rsp),%%rax\n", STACK_SPACE + 8 );
1061     output( "\tmovq %%rax,0xf8(%%rsp)\n" );
1062
1063     output( "\tstmxcsr 0x34(%%rsp)\n" );
1064     output( "\tfxsave 0x100(%%rsp)\n" );
1065     for (i = 0; i < 16; i++)
1066     {
1067         output( "\tmovdqa %%xmm%u,0x%x(%%rsp)\n", i, 0x1a0 + 16 * i );
1068         output_cfi( ".cfi_rel_offset %%xmm%u,0x%x", i, 0x1a0 + 16 * i );
1069     }
1070
1071     output( "\tmovw %%cs,0x38(%%rsp)\n" );
1072     output( "\tmovw %%ds,0x3a(%%rsp)\n" );
1073     output( "\tmovw %%es,0x3c(%%rsp)\n" );
1074     output( "\tmovw %%fs,0x3e(%%rsp)\n" );
1075     output( "\tmovw %%gs,0x40(%%rsp)\n" );
1076     output( "\tmovw %%ss,0x42(%%rsp)\n" );
1077     output( "\tpushfq\n" );
1078     output( "\tpopq %%rax\n" );
1079     output( "\tmovl %%eax,0x44(%%rsp)\n" );
1080
1081     output( "\tmovl $0x%x,0x30(%%rsp)\n", 0x0010000f );
1082
1083     /* transfer the arguments */
1084
1085     output( "\tmovq %%r8,%u(%%rsp)\n", STACK_SPACE + 32 );
1086     output( "\tmovq %%r9,%u(%%rsp)\n", STACK_SPACE + 40 );
1087     output( "\tmovq $4,%%rax\n" );
1088     output( "\tleaq %u(%%rsp),%%rsi\n", STACK_SPACE + 16 );
1089     output( "\tcmpq %%rax,%%rcx\n" );
1090     output( "\tcmovgq %%rcx,%%rax\n" );
1091     output( "\tmovq %%rsp,%%rbx\n" );
1092     output_cfi( ".cfi_def_cfa_register %%rbx" );
1093     output( "\tleaq 16(,%%rax,8),%%rax\n" );  /* add 8 for context arg and 8 for rounding */
1094     output( "\tandq $~15,%%rax\n" );
1095     output( "\tsubq %%rax,%%rsp\n" );
1096     output( "\tmovq %%rsp,%%rdi\n" );
1097     output( "\tjrcxz 1f\n" );
1098     output( "\tcld\n" );
1099     output( "\trep\n\tmovsq\n" );
1100     output( "1:\tmovq %%rbx,0(%%rdi)\n" );  /* context arg */
1101
1102     /* call the entry point */
1103
1104     output( "\tmovq %%rdx,%%rax\n" );
1105     output( "\tmovq 0(%%rsp),%%rcx\n" );
1106     output( "\tmovq 8(%%rsp),%%rdx\n" );
1107     output( "\tmovq 16(%%rsp),%%r8\n" );
1108     output( "\tmovq 24(%%rsp),%%r9\n" );
1109     output( "\tcallq *%%rax\n" );
1110
1111     /* restore the context structure */
1112
1113     output( "1:\tmovq 0x80(%%rbx),%%rcx\n" );
1114     output_cfi( ".cfi_same_value %%rcx" );
1115     output( "\tmovq 0x88(%%rbx),%%rdx\n" );
1116     output_cfi( ".cfi_same_value %%rdx" );
1117     output( "\tmovq 0xa0(%%rbx),%%rbp\n" );
1118     output_cfi( ".cfi_same_value %%rbp" );
1119     output( "\tmovq 0xa8(%%rbx),%%rsi\n" );
1120     output_cfi( ".cfi_same_value %%rsi" );
1121     output( "\tmovq 0xb0(%%rbx),%%rdi\n" );
1122     output_cfi( ".cfi_same_value %%rdi" );
1123     output( "\tmovq 0xb8(%%rbx),%%r8\n" );
1124     output_cfi( ".cfi_same_value %%r8" );
1125     output( "\tmovq 0xc0(%%rbx),%%r9\n" );
1126     output_cfi( ".cfi_same_value %%r9" );
1127     output( "\tmovq 0xc8(%%rbx),%%r10\n" );
1128     output_cfi( ".cfi_same_value %%r10" );
1129     output( "\tmovq 0xd0(%%rbx),%%r11\n" );
1130     output_cfi( ".cfi_same_value %%r11" );
1131     output( "\tmovq 0xd8(%%rbx),%%r12\n" );
1132     output_cfi( ".cfi_same_value %%r12" );
1133     output( "\tmovq 0xe0(%%rbx),%%r13\n" );
1134     output_cfi( ".cfi_same_value %%r13" );
1135     output( "\tmovq 0xe8(%%rbx),%%r14\n" );
1136     output_cfi( ".cfi_same_value %%r14" );
1137     output( "\tmovq 0xf0(%%rbx),%%r15\n" );
1138     output_cfi( ".cfi_same_value %%r15" );
1139
1140     for (i = 0; i < 16; i++)
1141     {
1142         output( "\tmovdqa 0x%x(%%rbx),%%xmm%u\n", 0x1a0 + 16 * i, i );
1143         output_cfi( ".cfi_same_value %%xmm%u", i );
1144     }
1145     output( "\tfxrstor 0x100(%%rbx)\n" );
1146     output( "\tldmxcsr 0x34(%%rbx)\n" );
1147
1148     output( "\tmovq 0xf8(%%rbx),%%rax\n" );  /* rip */
1149     output( "\tmovq %%rax,0(%%rsp)\n" );
1150     output( "\tmovw 0x38(%%rbx),%%ax\n" );   /* cs */
1151     output( "\tmovq %%rax,0x8(%%rsp)\n" );
1152     output( "\tmovl 0x44(%%rbx),%%eax\n" );  /* flags */
1153     output( "\tmovq %%rax,0x10(%%rsp)\n" );
1154     output( "\tmovq 0x98(%%rbx),%%rax\n" );  /* rsp */
1155     output( "\tmovq %%rax,0x18(%%rsp)\n" );
1156     output( "\tmovw 0x42(%%rbx),%%ax\n" );   /* ss */
1157     output( "\tmovq %%rax,0x20(%%rsp)\n" );
1158     output( "\tmovq 0x78(%%rbx),%%rax\n" );
1159     output( "\tmovq 0x90(%%rbx),%%rbx\n" );
1160     output( "\tiretq\n" );
1161     output_cfi( ".cfi_endproc" );
1162
1163     output_function_size( "__wine_call_from_regs" );
1164
1165     function_header( "__wine_restore_regs" );
1166     output_cfi( ".cfi_startproc" );
1167     output( "\tmovq %%rcx,%%rbx\n" );
1168     output( "\tjmp 1b\n" );
1169     output_cfi( ".cfi_endproc" );
1170     output_function_size( "__wine_restore_regs" );
1171 }
1172
1173
1174 /*******************************************************************
1175  *         output_asm_relays
1176  *
1177  * Build all the assembly relay callbacks
1178  */
1179 void output_asm_relays(void)
1180 {
1181     switch (target_cpu)
1182     {
1183     case CPU_x86:
1184         build_call_from_regs_x86();
1185         break;
1186     case CPU_x86_64:
1187         build_call_from_regs_x86_64();
1188         break;
1189     default:
1190         break;
1191     }
1192 }