libffi/src/x86/ffi64.c

   1 /* -----------------------------------------------------------------------
   2    ffi.c - Copyright (c) 2002  Bo Thorsen <bo@suse.de>
   3
   4    x86-64 Foreign Function Interface
   5
   6    Permission is hereby granted, free of charge, to any person obtaining
   7    a copy of this software and associated documentation files (the
   8    ``Software''), to deal in the Software without restriction, including
   9    without limitation the rights to use, copy, modify, merge, publish,
  10    distribute, sublicense, and/or sell copies of the Software, and to
  11    permit persons to whom the Software is furnished to do so, subject to
  12    the following conditions:
  13
  14    The above copyright notice and this permission notice shall be included
  15    in all copies or substantial portions of the Software.
  16
  17    THE SOFTWARE IS PROVIDED ``AS IS'', WITHOUT WARRANTY OF ANY KIND, EXPRESS
  18    OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  19    MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
  20    IN NO EVENT SHALL CYGNUS SOLUTIONS BE LIABLE FOR ANY CLAIM, DAMAGES OR
  21    OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  22    ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  23    OTHER DEALINGS IN THE SOFTWARE.
  24    ----------------------------------------------------------------------- */
  25
  26 #include <ffi.h>
  27 #include <ffi_common.h>
  28
  29 #include <stdlib.h>
  30 #include <stdarg.h>
  31
  32 #ifdef __x86_64__
  33
  34 #define MAX_GPR_REGS 6
  35 #define MAX_SSE_REGS 8
  36
  37 struct register_args
  38 {
  39   /* Registers for argument passing.  */
  40   UINT64 gpr[MAX_GPR_REGS];
  41   __int128_t sse[MAX_SSE_REGS];
  42 };
  43
  44 extern void ffi_call_unix64 (void *args, unsigned long bytes, unsigned flags,
  45                              void *raddr, void (*fnaddr)());
  46
  47 /* All reference to register classes here is identical to the code in
  48    gcc/config/i386/i386.c. Do *not* change one without the other.  */
  49
  50 /* Register class used for passing given 64bit part of the argument.
  51    These represent classes as documented by the PS ABI, with the exception
  52    of SSESF, SSEDF classes, that are basically SSE class, just gcc will
  53    use SF or DFmode move instead of DImode to avoid reformating penalties.
  54
  55    Similary we play games with INTEGERSI_CLASS to use cheaper SImode moves
  56    whenever possible (upper half does contain padding).  */
  57 enum x86_64_reg_class
  58   {
  59     X86_64_NO_CLASS,
  60     X86_64_INTEGER_CLASS,
  61     X86_64_INTEGERSI_CLASS,
  62     X86_64_SSE_CLASS,
  63     X86_64_SSESF_CLASS,
  64     X86_64_SSEDF_CLASS,
  65     X86_64_SSEUP_CLASS,
  66     X86_64_X87_CLASS,
  67     X86_64_X87UP_CLASS,
  68     X86_64_COMPLEX_X87_CLASS,
  69     X86_64_MEMORY_CLASS
  70   };
  71
  72 #define MAX_CLASSES 4
  73
  74 #define SSE_CLASS_P(X)  ((X) >= X86_64_SSE_CLASS && X <= X86_64_SSEUP_CLASS)
  75
  76 /* x86-64 register passing implementation.  See x86-64 ABI for details.  Goal
  77    of this code is to classify each 8bytes of incoming argument by the register
  78    class and assign registers accordingly.  */
  79
  80 /* Return the union class of CLASS1 and CLASS2.
  81    See the x86-64 PS ABI for details.  */
  82
  83 static enum x86_64_reg_class
  84 merge_classes (enum x86_64_reg_class class1, enum x86_64_reg_class class2)
  85 {
  86   /* Rule #1: If both classes are equal, this is the resulting class.  */
  87   if (class1 == class2)
  88     return class1;
  89
  90   /* Rule #2: If one of the classes is NO_CLASS, the resulting class is
  91      the other class.  */
  92   if (class1 == X86_64_NO_CLASS)
  93     return class2;
  94   if (class2 == X86_64_NO_CLASS)
  95     return class1;
  96
  97   /* Rule #3: If one of the classes is MEMORY, the result is MEMORY.  */
  98   if (class1 == X86_64_MEMORY_CLASS || class2 == X86_64_MEMORY_CLASS)
  99     return X86_64_MEMORY_CLASS;
 100
 101   /* Rule #4: If one of the classes is INTEGER, the result is INTEGER.  */
 102   if ((class1 == X86_64_INTEGERSI_CLASS && class2 == X86_64_SSESF_CLASS)
 103       || (class2 == X86_64_INTEGERSI_CLASS && class1 == X86_64_SSESF_CLASS))
 104     return X86_64_INTEGERSI_CLASS;
 105   if (class1 == X86_64_INTEGER_CLASS || class1 == X86_64_INTEGERSI_CLASS
 106       || class2 == X86_64_INTEGER_CLASS || class2 == X86_64_INTEGERSI_CLASS)
 107     return X86_64_INTEGER_CLASS;
 108
 109   /* Rule #5: If one of the classes is X87, X87UP, or COMPLEX_X87 class,
 110      MEMORY is used.  */
 111   if (class1 == X86_64_X87_CLASS
 112       || class1 == X86_64_X87UP_CLASS
 113       || class1 == X86_64_COMPLEX_X87_CLASS
 114       || class2 == X86_64_X87_CLASS
 115       || class2 == X86_64_X87UP_CLASS
 116       || class2 == X86_64_COMPLEX_X87_CLASS)
 117     return X86_64_MEMORY_CLASS;
 118
 119   /* Rule #6: Otherwise class SSE is used.  */
 120   return X86_64_SSE_CLASS;
 121 }
 122
 123 /* Classify the argument of type TYPE and mode MODE.
 124    CLASSES will be filled by the register class used to pass each word
 125    of the operand.  The number of words is returned.  In case the parameter
 126    should be passed in memory, 0 is returned. As a special case for zero
 127    sized containers, classes[0] will be NO_CLASS and 1 is returned.
 128
 129    See the x86-64 PS ABI for details.
 130 */
 131 static int
 132 classify_argument (ffi_type *type, enum x86_64_reg_class classes[],
 133                    size_t byte_offset)
 134 {
 135   switch (type->type)
 136     {
 137     case FFI_TYPE_UINT8:
 138     case FFI_TYPE_SINT8:
 139     case FFI_TYPE_UINT16:
 140     case FFI_TYPE_SINT16:
 141     case FFI_TYPE_UINT32:
 142     case FFI_TYPE_SINT32:
 143     case FFI_TYPE_UINT64:
 144     case FFI_TYPE_SINT64:
 145     case FFI_TYPE_POINTER:
 146       if (byte_offset + type->size <= 4)
 147         classes[0] = X86_64_INTEGERSI_CLASS;
 148       else
 149         classes[0] = X86_64_INTEGER_CLASS;
 150       return 1;
 151     case FFI_TYPE_FLOAT:
 152       if (byte_offset == 0)
 153         classes[0] = X86_64_SSESF_CLASS;
 154       else
 155         classes[0] = X86_64_SSE_CLASS;
 156       return 1;
 157     case FFI_TYPE_DOUBLE:
 158       classes[0] = X86_64_SSEDF_CLASS;
 159       return 1;
 160     case FFI_TYPE_LONGDOUBLE:
 161       classes[0] = X86_64_X87_CLASS;
 162       classes[1] = X86_64_X87UP_CLASS;
 163       return 2;
 164     case FFI_TYPE_STRUCT:
 165       {
 166         const int UNITS_PER_WORD = 8;
 167         int words = (type->size + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
 168         ffi_type **ptr;
 169         int i;
 170         enum x86_64_reg_class subclasses[MAX_CLASSES];
 171
 172         /* If the struct is larger than 16 bytes, pass it on the stack.  */
 173         if (type->size > 16)
 174           return 0;
 175
 176         for (i = 0; i < words; i++)
 177           classes[i] = X86_64_NO_CLASS;
 178
 179         /* Merge the fields of structure.  */
 180         for (ptr = type->elements; *ptr != NULL; ptr++)
 181           {
 182             int num;
 183
 184             byte_offset = ALIGN (byte_offset, (*ptr)->alignment);
 185
 186             num = classify_argument (*ptr, subclasses, byte_offset % 8);
 187             if (num == 0)
 188               return 0;
 189             for (i = 0; i < num; i++)
 190               {
 191                 int pos = byte_offset / 8;
 192                 classes[i + pos] =
 193                   merge_classes (subclasses[i], classes[i + pos]);
 194               }
 195
 196             byte_offset += (*ptr)->size;
 197           }
 198
 199         /* Final merger cleanup.  */
 200         for (i = 0; i < words; i++)
 201           {
 202             /* If one class is MEMORY, everything should be passed in
 203                memory.  */
 204             if (classes[i] == X86_64_MEMORY_CLASS)
 205               return 0;
 206
 207             /* The X86_64_SSEUP_CLASS should be always preceded by
 208                X86_64_SSE_CLASS.  */
 209             if (classes[i] == X86_64_SSEUP_CLASS
 210                 && (i == 0 || classes[i - 1] != X86_64_SSE_CLASS))
 211               classes[i] = X86_64_SSE_CLASS;
 212
 213             /*  X86_64_X87UP_CLASS should be preceded by X86_64_X87_CLASS.  */
 214             if (classes[i] == X86_64_X87UP_CLASS
 215                 && (i == 0 || classes[i - 1] != X86_64_X87_CLASS))
 216               classes[i] = X86_64_SSE_CLASS;
 217           }
 218         return words;
 219       }
 220
 221     default:
 222       FFI_ASSERT(0);
 223     }
 224   return 0; /* Never reached.  */
 225 }
 226
 227 /* Examine the argument and return set number of register required in each
 228    class.  Return zero iff parameter should be passed in memory, otherwise
 229    the number of registers.  */
 230
 231 static int
 232 examine_argument (ffi_type *type, enum x86_64_reg_class classes[MAX_CLASSES],
 233                   _Bool in_return, int *pngpr, int *pnsse)
 234 {
 235   int i, n, ngpr, nsse;
 236
 237   n = classify_argument (type, classes, 0);
 238   if (n == 0)
 239     return 0;
 240
 241   ngpr = nsse = 0;
 242   for (i = 0; i < n; ++i)
 243     switch (classes[i])
 244       {
 245       case X86_64_INTEGER_CLASS:
 246       case X86_64_INTEGERSI_CLASS:
 247         ngpr++;
 248         break;
 249       case X86_64_SSE_CLASS:
 250       case X86_64_SSESF_CLASS:
 251       case X86_64_SSEDF_CLASS:
 252         nsse++;
 253         break;
 254       case X86_64_NO_CLASS:
 255       case X86_64_SSEUP_CLASS:
 256         break;
 257       case X86_64_X87_CLASS:
 258       case X86_64_X87UP_CLASS:
 259       case X86_64_COMPLEX_X87_CLASS:
 260         return in_return != 0;
 261       default:
 262         abort ();
 263       }
 264
 265   *pngpr = ngpr;
 266   *pnsse = nsse;
 267
 268   return n;
 269 }
 270
 271 /* Perform machine dependent cif processing.  */
 272
 273 ffi_status
 274 ffi_prep_cif_machdep (ffi_cif *cif)
 275 {
 276   int gprcount, ssecount, i, avn, n, ngpr, nsse, flags;
 277   enum x86_64_reg_class classes[MAX_CLASSES];
 278   size_t bytes;
 279
 280   gprcount = ssecount = 0;
 281
 282   flags = cif->rtype->type;
 283   if (flags != FFI_TYPE_VOID)
 284     {
 285       n = examine_argument (cif->rtype, classes, 1, &ngpr, &nsse);
 286       if (n == 0)
 287         {
 288           /* The return value is passed in memory.  A pointer to that
 289              memory is the first argument.  Allocate a register for it.  */
 290           gprcount++;
 291           /* We don't have to do anything in asm for the return.  */
 292           flags = FFI_TYPE_VOID;
 293         }
 294       else if (flags == FFI_TYPE_STRUCT)
 295         {
 296           /* Mark which registers the result appears in.  */
 297           _Bool sse0 = SSE_CLASS_P (classes[0]);
 298           _Bool sse1 = n == 2 && SSE_CLASS_P (classes[1]);
 299           if (sse0 && !sse1)
 300             flags |= 1 << 8;
 301           else if (!sse0 && sse1)
 302             flags |= 1 << 9;
 303           else if (sse0 && sse1)
 304             flags |= 1 << 10;
 305           /* Mark the true size of the structure.  */
 306           flags |= cif->rtype->size << 11;
 307         }
 308     }
 309   cif->flags = flags;
 310
 311   /* Go over all arguments and determine the way they should be passed.
 312      If it's in a register and there is space for it, let that be so. If
 313      not, add it's size to the stack byte count.  */
 314   for (bytes = 0, i = 0, avn = cif->nargs; i < avn; i++)
 315     {
 316       if (examine_argument (cif->arg_types[i], classes, 0, &ngpr, &nsse) == 0
 317           || gprcount + ngpr > MAX_GPR_REGS
 318           || ssecount + nsse > MAX_SSE_REGS)
 319         {
 320           long align = cif->arg_types[i]->alignment;
 321
 322           if (align < 8)
 323             align = 8;
 324
 325           bytes = ALIGN(bytes, align);
 326           bytes += cif->arg_types[i]->size;
 327         }
 328       else
 329         {
 330           gprcount += ngpr;
 331           ssecount += nsse;
 332         }
 333     }
 334   cif->bytes = bytes;
 335
 336   return FFI_OK;
 337 }
 338
 339 void
 340 ffi_call (ffi_cif *cif, void (*fn)(), void *rvalue, void **avalue)
 341 {
 342   enum x86_64_reg_class classes[MAX_CLASSES];
 343   char *stack, *argp;
 344   ffi_type **arg_types;
 345   int gprcount, ssecount, ngpr, nsse, i, avn;
 346   _Bool ret_in_memory;
 347   struct register_args *reg_args;
 348
 349   /* Can't call 32-bit mode from 64-bit mode.  */
 350   FFI_ASSERT (cif->abi == FFI_UNIX64);
 351
 352   /* If the return value is a struct and we don't have a return value
 353      address then we need to make one.  Note the setting of flags to
 354      VOID above in ffi_prep_cif_machdep.  */
 355   ret_in_memory = (cif->rtype->type == FFI_TYPE_STRUCT
 356                    && cif->flags == FFI_TYPE_VOID);
 357   if (rvalue == NULL && ret_in_memory)
 358     rvalue = alloca (cif->rtype->size);
 359
 360   /* Allocate the space for the arguments, plus 4 words of temp space.  */
 361   stack = alloca (sizeof (struct register_args) + cif->bytes + 4*8);
 362   reg_args = (struct register_args *) stack;
 363   argp = stack + sizeof (struct register_args);
 364
 365   gprcount = ssecount = 0;
 366
 367   /* If the return value is passed in memory, add the pointer as the
 368      first integer argument.  */
 369   if (ret_in_memory)
 370     reg_args->gpr[gprcount++] = (long) rvalue;
 371
 372   avn = cif->nargs;
 373   arg_types = cif->arg_types;
 374
 375   for (i = 0; i < avn; ++i)
 376     {
 377       size_t size = arg_types[i]->size;
 378       int n;
 379
 380       n = examine_argument (arg_types[i], classes, 0, &ngpr, &nsse);
 381       if (n == 0
 382           || gprcount + ngpr > MAX_GPR_REGS
 383           || ssecount + nsse > MAX_SSE_REGS)
 384         {
 385           long align = arg_types[i]->alignment;
 386
 387           /* Stack arguments are *always* at least 8 byte aligned.  */
 388           if (align < 8)
 389             align = 8;
 390
 391           /* Pass this argument in memory.  */
 392           argp = (void *) ALIGN (argp, align);
 393           memcpy (argp, avalue[i], size);
 394           argp += size;
 395         }
 396       else
 397         {
 398           /* The argument is passed entirely in registers.  */
 399           char *a = (char *) avalue[i];
 400           int j;
 401
 402           for (j = 0; j < n; j++, a += 8, size -= 8)
 403             {
 404               switch (classes[j])
 405                 {
 406                 case X86_64_INTEGER_CLASS:
 407                 case X86_64_INTEGERSI_CLASS:
 408                   reg_args->gpr[gprcount] = 0;
 409                   memcpy (&reg_args->gpr[gprcount], a, size < 8 ? size : 8);
 410                   gprcount++;
 411                   break;
 412                 case X86_64_SSE_CLASS:
 413                 case X86_64_SSEDF_CLASS:
 414                   reg_args->sse[ssecount++] = *(UINT64 *) a;
 415                   break;
 416                 case X86_64_SSESF_CLASS:
 417                   reg_args->sse[ssecount++] = *(UINT32 *) a;
 418                   break;
 419                 default:
 420                   abort();
 421                 }
 422             }
 423         }
 424     }
 425
 426   ffi_call_unix64 (stack, cif->bytes + sizeof (struct register_args),
 427                    cif->flags, rvalue, fn);
 428 }
 429
 430
 431 extern void ffi_closure_unix64(void);
 432
 433 ffi_status
 434 ffi_prep_closure (ffi_closure* closure,
 435                   ffi_cif* cif,
 436                   void (*fun)(ffi_cif*, void*, void**, void*),
 437                   void *user_data)
 438 {
 439   volatile unsigned short *tramp;
 440
 441   tramp = (volatile unsigned short *) &closure->tramp[0];
 442   tramp[0] = 0xbb49;            /* mov <code>, %r11     */
 443   tramp[5] = 0xba49;            /* mov <data>, %r10     */
 444   tramp[10] = 0xff49;           /* jmp *%r11    */
 445   tramp[11] = 0x00e3;
 446   *(void * volatile *) &tramp[1] = ffi_closure_unix64;
 447   *(void * volatile *) &tramp[6] = closure;
 448
 449   closure->cif = cif;
 450   closure->fun = fun;
 451   closure->user_data = user_data;
 452
 453   return FFI_OK;
 454 }
 455
 456 int
 457 ffi_closure_unix64_inner(ffi_closure *closure, void *rvalue,
 458                          struct register_args *reg_args, char *argp)
 459 {
 460   ffi_cif *cif;
 461   void **avalue;
 462   ffi_type **arg_types;
 463   long i, avn;
 464   int gprcount, ssecount, ngpr, nsse;
 465   int ret;
 466
 467   cif = closure->cif;
 468   avalue = alloca(cif->nargs * sizeof(void *));
 469   gprcount = ssecount = 0;
 470
 471   ret = cif->rtype->type;
 472   if (ret != FFI_TYPE_VOID)
 473     {
 474       enum x86_64_reg_class classes[MAX_CLASSES];
 475       int n = examine_argument (cif->rtype, classes, 1, &ngpr, &nsse);
 476       if (n == 0)
 477         {
 478           /* The return value goes in memory.  Arrange for the closure
 479              return value to go directly back to the original caller.  */
 480           rvalue = (void *) reg_args->gpr[gprcount++];
 481           /* We don't have to do anything in asm for the return.  */
 482           ret = FFI_TYPE_VOID;
 483         }
 484       else if (ret == FFI_TYPE_STRUCT && n == 2)
 485         {
 486           /* Mark which register the second word of the structure goes in.  */
 487           _Bool sse0 = SSE_CLASS_P (classes[0]);
 488           _Bool sse1 = SSE_CLASS_P (classes[1]);
 489           if (!sse0 && sse1)
 490             ret |= 1 << 8;
 491           else if (sse0 && !sse1)
 492             ret |= 1 << 9;
 493         }
 494     }
 495
 496   avn = cif->nargs;
 497   arg_types = cif->arg_types;
 498
 499   for (i = 0; i < avn; ++i)
 500     {
 501       enum x86_64_reg_class classes[MAX_CLASSES];
 502       int n;
 503
 504       n = examine_argument (arg_types[i], classes, 0, &ngpr, &nsse);
 505       if (n == 0
 506           || gprcount + ngpr > MAX_GPR_REGS
 507           || ssecount + nsse > MAX_SSE_REGS)
 508         {
 509           long align = arg_types[i]->alignment;
 510
 511           /* Stack arguments are *always* at least 8 byte aligned.  */
 512           if (align < 8)
 513             align = 8;
 514
 515           /* Pass this argument in memory.  */
 516           argp = (void *) ALIGN (argp, align);
 517           avalue[i] = argp;
 518           argp += arg_types[i]->size;
 519         }
 520       /* If the argument is in a single register, or two consecutive
 521          registers, then we can use that address directly.  */
 522       else if (n == 1
 523                || (n == 2
 524                    && SSE_CLASS_P (classes[0]) == SSE_CLASS_P (classes[1])))
 525         {
 526           /* The argument is in a single register.  */
 527           if (SSE_CLASS_P (classes[0]))
 528             {
 529               avalue[i] = &reg_args->sse[ssecount];
 530               ssecount += n;
 531             }
 532           else
 533             {
 534               avalue[i] = &reg_args->gpr[gprcount];
 535               gprcount += n;
 536             }
 537         }
 538       /* Otherwise, allocate space to make them consecutive.  */
 539       else
 540         {
 541           char *a = alloca (16);
 542           int j;
 543
 544           avalue[i] = a;
 545           for (j = 0; j < n; j++, a += 8)
 546             {
 547               if (SSE_CLASS_P (classes[j]))
 548                 memcpy (a, &reg_args->sse[ssecount++], 8);
 549               else
 550                 memcpy (a, &reg_args->gpr[gprcount++], 8);
 551             }
 552         }
 553     }
 554
 555   /* Invoke the closure.  */
 556   closure->fun (cif, rvalue, avalue, closure->user_data);
 557
 558   /* Tell assembly how to perform return type promotions.  */
 559   return ret;
 560 }
 561
 562 #endif /* __x86_64__ */