src/mesa/x86/rtasm/x86sse.c

   1 #if defined(USE_X86_ASM) || defined(SLANG_X86)
   2
   3 #include "imports.h"
   4 #include "x86sse.h"
   5
   6 #define DISASSEM 0
   7 #define X86_TWOB 0x0f
   8
   9 /* Emit bytes to the instruction stream:
  10  */
  11 static void emit_1b( struct x86_function *p, GLbyte b0 )
  12 {
  13    *(GLbyte *)(p->csr++) = b0;
  14 }
  15
  16 static void emit_1i( struct x86_function *p, GLint i0 )
  17 {
  18    *(GLint *)(p->csr) = i0;
  19    p->csr += 4;
  20 }
  21
  22 static void disassem( struct x86_function *p, const char *fn )
  23 {
  24 #if DISASSEM && 0
  25    if (fn && fn != p->fn) {
  26       _mesa_printf("0x%x: %s\n", p->csr, fn);
  27       p->fn = fn;
  28    }
  29 #endif
  30 }
  31
  32 static void emit_1ub_fn( struct x86_function *p, GLubyte b0, const char *fn )
  33 {
  34    disassem(p, fn);
  35    *(p->csr++) = b0;
  36 }
  37
  38 static void emit_2ub_fn( struct x86_function *p, GLubyte b0, GLubyte b1, const char *fn )
  39 {
  40    disassem(p, fn);
  41    *(p->csr++) = b0;
  42    *(p->csr++) = b1;
  43 }
  44
  45 static void emit_3ub_fn( struct x86_function *p, GLubyte b0, GLubyte b1, GLubyte b2, const char *fn )
  46 {
  47    disassem(p, fn);
  48    *(p->csr++) = b0;
  49    *(p->csr++) = b1;
  50    *(p->csr++) = b2;
  51 }
  52
  53 #define emit_1ub(p, b0)         emit_1ub_fn(p, b0, __FUNCTION__)
  54 #define emit_2ub(p, b0, b1)     emit_2ub_fn(p, b0, b1, __FUNCTION__)
  55 #define emit_3ub(p, b0, b1, b2) emit_3ub_fn(p, b0, b1, b2, __FUNCTION__)
  56
  57
  58
  59 /* Build a modRM byte + possible displacement.  No treatment of SIB
  60  * indexing.  BZZT - no way to encode an absolute address.
  61  */
  62 static void emit_modrm( struct x86_function *p,
  63                         struct x86_reg reg,
  64                         struct x86_reg regmem )
  65 {
  66    GLubyte val = 0;
  67
  68    assert(reg.mod == mod_REG);
  69
  70    val |= regmem.mod << 6;      /* mod field */
  71    val |= reg.idx << 3;         /* reg field */
  72    val |= regmem.idx;           /* r/m field */
  73
  74    emit_1ub_fn(p, val, 0);
  75
  76    /* Oh-oh we've stumbled into the SIB thing.
  77     */
  78    if (regmem.file == file_REG32 &&
  79        regmem.idx == reg_SP) {
  80       emit_1ub_fn(p, 0x24, 0);          /* simplistic! */
  81    }
  82
  83    switch (regmem.mod) {
  84    case mod_REG:
  85    case mod_INDIRECT:
  86       break;
  87    case mod_DISP8:
  88       emit_1b(p, regmem.disp);
  89       break;
  90    case mod_DISP32:
  91       emit_1i(p, regmem.disp);
  92       break;
  93    default:
  94       assert(0);
  95       break;
  96    }
  97 }
  98
  99
 100 static void emit_modrm_noreg( struct x86_function *p,
 101                               GLuint op,
 102                               struct x86_reg regmem )
 103 {
 104    struct x86_reg dummy = x86_make_reg(file_REG32, op);
 105    emit_modrm(p, dummy, regmem);
 106 }
 107
 108 /* Many x86 instructions have two opcodes to cope with the situations
 109  * where the destination is a register or memory reference
 110  * respectively.  This function selects the correct opcode based on
 111  * the arguments presented.
 112  */
 113 static void emit_op_modrm( struct x86_function *p,
 114                            GLubyte op_dst_is_reg,
 115                            GLubyte op_dst_is_mem,
 116                            struct x86_reg dst,
 117                            struct x86_reg src )
 118 {
 119    switch (dst.mod) {
 120    case mod_REG:
 121       emit_1ub_fn(p, op_dst_is_reg, 0);
 122       emit_modrm(p, dst, src);
 123       break;
 124    case mod_INDIRECT:
 125    case mod_DISP32:
 126    case mod_DISP8:
 127       assert(src.mod == mod_REG);
 128       emit_1ub_fn(p, op_dst_is_mem, 0);
 129       emit_modrm(p, src, dst);
 130       break;
 131    default:
 132       assert(0);
 133       break;
 134    }
 135 }
 136
 137
 138
 139
 140
 141
 142
 143 /* Create and manipulate registers and regmem values:
 144  */
 145 struct x86_reg x86_make_reg( enum x86_reg_file file,
 146                              enum x86_reg_name idx )
 147 {
 148    struct x86_reg reg;
 149
 150    reg.file = file;
 151    reg.idx = idx;
 152    reg.mod = mod_REG;
 153    reg.disp = 0;
 154
 155    return reg;
 156 }
 157
 158 struct x86_reg x86_make_disp( struct x86_reg reg,
 159                               GLint disp )
 160 {
 161    assert(reg.file == file_REG32);
 162
 163    if (reg.mod == mod_REG)
 164       reg.disp = disp;
 165    else
 166       reg.disp += disp;
 167
 168    if (reg.disp == 0)
 169       reg.mod = mod_INDIRECT;
 170    else if (reg.disp <= 127 && reg.disp >= -128)
 171       reg.mod = mod_DISP8;
 172    else
 173       reg.mod = mod_DISP32;
 174
 175    return reg;
 176 }
 177
 178 struct x86_reg x86_deref( struct x86_reg reg )
 179 {
 180    return x86_make_disp(reg, 0);
 181 }
 182
 183 struct x86_reg x86_get_base_reg( struct x86_reg reg )
 184 {
 185    return x86_make_reg( reg.file, reg.idx );
 186 }
 187
 188 GLubyte *x86_get_label( struct x86_function *p )
 189 {
 190    return p->csr;
 191 }
 192
 193
 194
 195 /***********************************************************************
 196  * x86 instructions
 197  */
 198
 199
 200 void x86_jcc( struct x86_function *p,
 201               enum x86_cc cc,
 202               GLubyte *label )
 203 {
 204    GLint offset = label - (x86_get_label(p) + 2);
 205
 206    if (offset <= 127 && offset >= -128) {
 207       emit_1ub(p, 0x70 + cc);
 208       emit_1b(p, (GLbyte) offset);
 209    }
 210    else {
 211       offset = label - (x86_get_label(p) + 6);
 212       emit_2ub(p, 0x0f, 0x80 + cc);
 213       emit_1i(p, offset);
 214    }
 215 }
 216
 217 /* Always use a 32bit offset for forward jumps:
 218  */
 219 GLubyte *x86_jcc_forward( struct x86_function *p,
 220                           enum x86_cc cc )
 221 {
 222    emit_2ub(p, 0x0f, 0x80 + cc);
 223    emit_1i(p, 0);
 224    return x86_get_label(p);
 225 }
 226
 227 GLubyte *x86_jmp_forward( struct x86_function *p)
 228 {
 229    emit_1ub(p, 0xe9);
 230    emit_1i(p, 0);
 231    return x86_get_label(p);
 232 }
 233
 234 GLubyte *x86_call_forward( struct x86_function *p)
 235 {
 236    emit_1ub(p, 0xe8);
 237    emit_1i(p, 0);
 238    return x86_get_label(p);
 239 }
 240
 241 /* Fixup offset from forward jump:
 242  */
 243 void x86_fixup_fwd_jump( struct x86_function *p,
 244                          GLubyte *fixup )
 245 {
 246    *(int *)(fixup - 4) = x86_get_label(p) - fixup;
 247 }
 248
 249 void x86_jmp( struct x86_function *p, GLubyte *label)
 250 {
 251    emit_1ub(p, 0xe9);
 252    emit_1i(p, label - x86_get_label(p) - 4);
 253 }
 254
 255 void x86_call( struct x86_function *p, GLubyte *label)
 256 {
 257    emit_1ub(p, 0xe8);
 258    emit_1i(p, label - x86_get_label(p) - 4);
 259 }
 260
 261 /* michal:
 262  * Temporary. As I need immediate operands, and dont want to mess with the codegen,
 263  * I load the immediate into general purpose register and use it.
 264  */
 265 void x86_mov_reg_imm( struct x86_function *p, struct x86_reg dst, GLint imm )
 266 {
 267    assert(dst.mod == mod_REG);
 268    emit_1ub(p, 0xb8 + dst.idx);
 269    emit_1i(p, imm);
 270 }
 271
 272 void x86_push( struct x86_function *p,
 273                struct x86_reg reg )
 274 {
 275    assert(reg.mod == mod_REG);
 276    emit_1ub(p, 0x50 + reg.idx);
 277    p->stack_offset += 4;
 278 }
 279
 280 void x86_pop( struct x86_function *p,
 281               struct x86_reg reg )
 282 {
 283    assert(reg.mod == mod_REG);
 284    emit_1ub(p, 0x58 + reg.idx);
 285    p->stack_offset -= 4;
 286 }
 287
 288 void x86_inc( struct x86_function *p,
 289               struct x86_reg reg )
 290 {
 291    assert(reg.mod == mod_REG);
 292    emit_1ub(p, 0x40 + reg.idx);
 293 }
 294
 295 void x86_dec( struct x86_function *p,
 296               struct x86_reg reg )
 297 {
 298    assert(reg.mod == mod_REG);
 299    emit_1ub(p, 0x48 + reg.idx);
 300 }
 301
 302 void x86_ret( struct x86_function *p )
 303 {
 304    emit_1ub(p, 0xc3);
 305 }
 306
 307 void x86_sahf( struct x86_function *p )
 308 {
 309    emit_1ub(p, 0x9e);
 310 }
 311
 312 void x86_mov( struct x86_function *p,
 313               struct x86_reg dst,
 314               struct x86_reg src )
 315 {
 316    emit_op_modrm( p, 0x8b, 0x89, dst, src );
 317 }
 318
 319 void x86_xor( struct x86_function *p,
 320               struct x86_reg dst,
 321               struct x86_reg src )
 322 {
 323    emit_op_modrm( p, 0x33, 0x31, dst, src );
 324 }
 325
 326 void x86_cmp( struct x86_function *p,
 327               struct x86_reg dst,
 328               struct x86_reg src )
 329 {
 330    emit_op_modrm( p, 0x3b, 0x39, dst, src );
 331 }
 332
 333 void x86_lea( struct x86_function *p,
 334               struct x86_reg dst,
 335               struct x86_reg src )
 336 {
 337    emit_1ub(p, 0x8d);
 338    emit_modrm( p, dst, src );
 339 }
 340
 341 void x86_test( struct x86_function *p,
 342                struct x86_reg dst,
 343                struct x86_reg src )
 344 {
 345    emit_1ub(p, 0x85);
 346    emit_modrm( p, dst, src );
 347 }
 348
 349 void x86_add( struct x86_function *p,
 350                struct x86_reg dst,
 351                struct x86_reg src )
 352 {
 353    emit_op_modrm(p, 0x03, 0x01, dst, src );
 354 }
 355
 356 void x86_mul( struct x86_function *p,
 357                struct x86_reg src )
 358 {
 359    assert (src.file == file_REG32 && src.mod == mod_REG);
 360    emit_op_modrm(p, 0xf7, 0, x86_make_reg (file_REG32, reg_SP), src );
 361 }
 362
 363 void x86_sub( struct x86_function *p,
 364                struct x86_reg dst,
 365                struct x86_reg src )
 366 {
 367    emit_op_modrm(p, 0x2b, 0x29, dst, src );
 368 }
 369
 370
 371
 372 /***********************************************************************
 373  * SSE instructions
 374  */
 375
 376
 377 void sse_movss( struct x86_function *p,
 378                 struct x86_reg dst,
 379                 struct x86_reg src )
 380 {
 381    emit_2ub(p, 0xF3, X86_TWOB);
 382    emit_op_modrm( p, 0x10, 0x11, dst, src );
 383 }
 384
 385 void sse_movaps( struct x86_function *p,
 386                  struct x86_reg dst,
 387                  struct x86_reg src )
 388 {
 389    emit_1ub(p, X86_TWOB);
 390    emit_op_modrm( p, 0x28, 0x29, dst, src );
 391 }
 392
 393 void sse_movups( struct x86_function *p,
 394                  struct x86_reg dst,
 395                  struct x86_reg src )
 396 {
 397    emit_1ub(p, X86_TWOB);
 398    emit_op_modrm( p, 0x10, 0x11, dst, src );
 399 }
 400
 401 void sse_movhps( struct x86_function *p,
 402                  struct x86_reg dst,
 403                  struct x86_reg src )
 404 {
 405    assert(dst.mod != mod_REG || src.mod != mod_REG);
 406    emit_1ub(p, X86_TWOB);
 407    emit_op_modrm( p, 0x16, 0x17, dst, src ); /* cf movlhps */
 408 }
 409
 410 void sse_movlps( struct x86_function *p,
 411                  struct x86_reg dst,
 412                  struct x86_reg src )
 413 {
 414    assert(dst.mod != mod_REG || src.mod != mod_REG);
 415    emit_1ub(p, X86_TWOB);
 416    emit_op_modrm( p, 0x12, 0x13, dst, src ); /* cf movhlps */
 417 }
 418
 419 void sse_maxps( struct x86_function *p,
 420                 struct x86_reg dst,
 421                 struct x86_reg src )
 422 {
 423    emit_2ub(p, X86_TWOB, 0x5F);
 424    emit_modrm( p, dst, src );
 425 }
 426
 427 void sse_divss( struct x86_function *p,
 428                 struct x86_reg dst,
 429                 struct x86_reg src )
 430 {
 431    emit_3ub(p, 0xF3, X86_TWOB, 0x5E);
 432    emit_modrm( p, dst, src );
 433 }
 434
 435 void sse_minps( struct x86_function *p,
 436                 struct x86_reg dst,
 437                 struct x86_reg src )
 438 {
 439    emit_2ub(p, X86_TWOB, 0x5D);
 440    emit_modrm( p, dst, src );
 441 }
 442
 443 void sse_subps( struct x86_function *p,
 444                 struct x86_reg dst,
 445                 struct x86_reg src )
 446 {
 447    emit_2ub(p, X86_TWOB, 0x5C);
 448    emit_modrm( p, dst, src );
 449 }
 450
 451 void sse_mulps( struct x86_function *p,
 452                 struct x86_reg dst,
 453                 struct x86_reg src )
 454 {
 455    emit_2ub(p, X86_TWOB, 0x59);
 456    emit_modrm( p, dst, src );
 457 }
 458
 459 void sse_addps( struct x86_function *p,
 460                 struct x86_reg dst,
 461                 struct x86_reg src )
 462 {
 463    emit_2ub(p, X86_TWOB, 0x58);
 464    emit_modrm( p, dst, src );
 465 }
 466
 467 void sse_addss( struct x86_function *p,
 468                 struct x86_reg dst,
 469                 struct x86_reg src )
 470 {
 471    emit_3ub(p, 0xF3, X86_TWOB, 0x58);
 472    emit_modrm( p, dst, src );
 473 }
 474
 475 void sse_andps( struct x86_function *p,
 476                 struct x86_reg dst,
 477                 struct x86_reg src )
 478 {
 479    emit_2ub(p, X86_TWOB, 0x54);
 480    emit_modrm( p, dst, src );
 481 }
 482
 483
 484 void sse_rsqrtss( struct x86_function *p,
 485                   struct x86_reg dst,
 486                   struct x86_reg src )
 487 {
 488    emit_3ub(p, 0xF3, X86_TWOB, 0x52);
 489    emit_modrm( p, dst, src );
 490
 491 }
 492
 493 void sse_movhlps( struct x86_function *p,
 494                   struct x86_reg dst,
 495                   struct x86_reg src )
 496 {
 497    assert(dst.mod == mod_REG && src.mod == mod_REG);
 498    emit_2ub(p, X86_TWOB, 0x12);
 499    emit_modrm( p, dst, src );
 500 }
 501
 502 void sse_movlhps( struct x86_function *p,
 503                   struct x86_reg dst,
 504                   struct x86_reg src )
 505 {
 506    assert(dst.mod == mod_REG && src.mod == mod_REG);
 507    emit_2ub(p, X86_TWOB, 0x16);
 508    emit_modrm( p, dst, src );
 509 }
 510
 511
 512 void sse_cvtps2pi( struct x86_function *p,
 513                    struct x86_reg dst,
 514                    struct x86_reg src )
 515 {
 516    assert(dst.file == file_MMX &&
 517           (src.file == file_XMM || src.mod != mod_REG));
 518
 519    p->need_emms = 1;
 520
 521    emit_2ub(p, X86_TWOB, 0x2d);
 522    emit_modrm( p, dst, src );
 523 }
 524
 525
 526 /* Shufps can also be used to implement a reduced swizzle when dest ==
 527  * arg0.
 528  */
 529 void sse_shufps( struct x86_function *p,
 530                  struct x86_reg dest,
 531                  struct x86_reg arg0,
 532                  GLubyte shuf)
 533 {
 534    emit_2ub(p, X86_TWOB, 0xC6);
 535    emit_modrm(p, dest, arg0);
 536    emit_1ub(p, shuf);
 537 }
 538
 539 void sse_cmpps( struct x86_function *p,
 540                 struct x86_reg dest,
 541                 struct x86_reg arg0,
 542                 GLubyte cc)
 543 {
 544    emit_2ub(p, X86_TWOB, 0xC2);
 545    emit_modrm(p, dest, arg0);
 546    emit_1ub(p, cc);
 547 }
 548
 549 /***********************************************************************
 550  * SSE2 instructions
 551  */
 552
 553 /**
 554  * Perform a reduced swizzle:
 555  */
 556 void sse2_pshufd( struct x86_function *p,
 557                   struct x86_reg dest,
 558                   struct x86_reg arg0,
 559                   GLubyte shuf)
 560 {
 561    emit_3ub(p, 0x66, X86_TWOB, 0x70);
 562    emit_modrm(p, dest, arg0);
 563    emit_1ub(p, shuf);
 564 }
 565
 566 void sse2_cvtps2dq( struct x86_function *p,
 567                     struct x86_reg dst,
 568                     struct x86_reg src )
 569 {
 570    emit_3ub(p, 0x66, X86_TWOB, 0x5B);
 571    emit_modrm( p, dst, src );
 572 }
 573
 574 void sse2_packssdw( struct x86_function *p,
 575                     struct x86_reg dst,
 576                     struct x86_reg src )
 577 {
 578    emit_3ub(p, 0x66, X86_TWOB, 0x6B);
 579    emit_modrm( p, dst, src );
 580 }
 581
 582 void sse2_packsswb( struct x86_function *p,
 583                     struct x86_reg dst,
 584                     struct x86_reg src )
 585 {
 586    emit_3ub(p, 0x66, X86_TWOB, 0x63);
 587    emit_modrm( p, dst, src );
 588 }
 589
 590 void sse2_packuswb( struct x86_function *p,
 591                     struct x86_reg dst,
 592                     struct x86_reg src )
 593 {
 594    emit_3ub(p, 0x66, X86_TWOB, 0x67);
 595    emit_modrm( p, dst, src );
 596 }
 597
 598 void sse2_rcpss( struct x86_function *p,
 599                 struct x86_reg dst,
 600                 struct x86_reg src )
 601 {
 602    emit_3ub(p, 0xF3, X86_TWOB, 0x53);
 603    emit_modrm( p, dst, src );
 604 }
 605
 606 void sse2_movd( struct x86_function *p,
 607                 struct x86_reg dst,
 608                 struct x86_reg src )
 609 {
 610    emit_2ub(p, 0x66, X86_TWOB);
 611    emit_op_modrm( p, 0x6e, 0x7e, dst, src );
 612 }
 613
 614
 615
 616
 617 /***********************************************************************
 618  * x87 instructions
 619  */
 620 void x87_fist( struct x86_function *p, struct x86_reg dst )
 621 {
 622    emit_1ub(p, 0xdb);
 623    emit_modrm_noreg(p, 2, dst);
 624 }
 625
 626 void x87_fistp( struct x86_function *p, struct x86_reg dst )
 627 {
 628    emit_1ub(p, 0xdb);
 629    emit_modrm_noreg(p, 3, dst);
 630 }
 631
 632 void x87_fldz( struct x86_function *p )
 633 {
 634    emit_2ub(p, 0xd9, 0xee);
 635 }
 636
 637
 638 void x87_fldcw( struct x86_function *p, struct x86_reg arg )
 639 {
 640    assert(arg.file == file_REG32);
 641    assert(arg.mod != mod_REG);
 642    emit_1ub(p, 0xd9);
 643    emit_modrm_noreg(p, 5, arg);
 644 }
 645
 646 void x87_fld1( struct x86_function *p )
 647 {
 648    emit_2ub(p, 0xd9, 0xe8);
 649 }
 650
 651 void x87_fldl2e( struct x86_function *p )
 652 {
 653    emit_2ub(p, 0xd9, 0xea);
 654 }
 655
 656 void x87_fldln2( struct x86_function *p )
 657 {
 658    emit_2ub(p, 0xd9, 0xed);
 659 }
 660
 661 void x87_fwait( struct x86_function *p )
 662 {
 663    emit_1ub(p, 0x9b);
 664 }
 665
 666 void x87_fnclex( struct x86_function *p )
 667 {
 668    emit_2ub(p, 0xdb, 0xe2);
 669 }
 670
 671 void x87_fclex( struct x86_function *p )
 672 {
 673    x87_fwait(p);
 674    x87_fnclex(p);
 675 }
 676
 677
 678 static void x87_arith_op( struct x86_function *p, struct x86_reg dst, struct x86_reg arg,
 679                           GLubyte dst0ub0,
 680                           GLubyte dst0ub1,
 681                           GLubyte arg0ub0,
 682                           GLubyte arg0ub1,
 683                           GLubyte argmem_noreg)
 684 {
 685    assert(dst.file == file_x87);
 686
 687    if (arg.file == file_x87) {
 688       if (dst.idx == 0)
 689          emit_2ub(p, dst0ub0, dst0ub1+arg.idx);
 690       else if (arg.idx == 0)
 691          emit_2ub(p, arg0ub0, arg0ub1+arg.idx);
 692       else
 693          assert(0);
 694    }
 695    else if (dst.idx == 0) {
 696       assert(arg.file = file_REG32);
 697       emit_1ub(p, 0xd8);
 698       emit_modrm_noreg(p, argmem_noreg, arg);
 699    }
 700    else
 701       assert(0);
 702 }
 703
 704 void x87_fmul( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 705 {
 706    x87_arith_op(p, dst, arg,
 707                 0xd8, 0xc8,
 708                 0xdc, 0xc8,
 709                 4);
 710 }
 711
 712 void x87_fsub( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 713 {
 714    x87_arith_op(p, dst, arg,
 715                 0xd8, 0xe0,
 716                 0xdc, 0xe8,
 717                 4);
 718 }
 719
 720 void x87_fsubr( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 721 {
 722    x87_arith_op(p, dst, arg,
 723                 0xd8, 0xe8,
 724                 0xdc, 0xe0,
 725                 5);
 726 }
 727
 728 void x87_fadd( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 729 {
 730    x87_arith_op(p, dst, arg,
 731                 0xd8, 0xc0,
 732                 0xdc, 0xc0,
 733                 0);
 734 }
 735
 736 void x87_fdiv( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 737 {
 738    x87_arith_op(p, dst, arg,
 739                 0xd8, 0xf0,
 740                 0xdc, 0xf8,
 741                 6);
 742 }
 743
 744 void x87_fdivr( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 745 {
 746    x87_arith_op(p, dst, arg,
 747                 0xd8, 0xf8,
 748                 0xdc, 0xf0,
 749                 7);
 750 }
 751
 752 void x87_fmulp( struct x86_function *p, struct x86_reg dst )
 753 {
 754    assert(dst.file == file_x87);
 755    assert(dst.idx >= 1);
 756    emit_2ub(p, 0xde, 0xc8+dst.idx);
 757 }
 758
 759 void x87_fsubp( struct x86_function *p, struct x86_reg dst )
 760 {
 761    assert(dst.file == file_x87);
 762    assert(dst.idx >= 1);
 763    emit_2ub(p, 0xde, 0xe8+dst.idx);
 764 }
 765
 766 void x87_fsubrp( struct x86_function *p, struct x86_reg dst )
 767 {
 768    assert(dst.file == file_x87);
 769    assert(dst.idx >= 1);
 770    emit_2ub(p, 0xde, 0xe0+dst.idx);
 771 }
 772
 773 void x87_faddp( struct x86_function *p, struct x86_reg dst )
 774 {
 775    assert(dst.file == file_x87);
 776    assert(dst.idx >= 1);
 777    emit_2ub(p, 0xde, 0xc0+dst.idx);
 778 }
 779
 780 void x87_fdivp( struct x86_function *p, struct x86_reg dst )
 781 {
 782    assert(dst.file == file_x87);
 783    assert(dst.idx >= 1);
 784    emit_2ub(p, 0xde, 0xf8+dst.idx);
 785 }
 786
 787 void x87_fdivrp( struct x86_function *p, struct x86_reg dst )
 788 {
 789    assert(dst.file == file_x87);
 790    assert(dst.idx >= 1);
 791    emit_2ub(p, 0xde, 0xf0+dst.idx);
 792 }
 793
 794 void x87_fucom( struct x86_function *p, struct x86_reg arg )
 795 {
 796    assert(arg.file == file_x87);
 797    emit_2ub(p, 0xdd, 0xe0+arg.idx);
 798 }
 799
 800 void x87_fucomp( struct x86_function *p, struct x86_reg arg )
 801 {
 802    assert(arg.file == file_x87);
 803    emit_2ub(p, 0xdd, 0xe8+arg.idx);
 804 }
 805
 806 void x87_fucompp( struct x86_function *p )
 807 {
 808    emit_2ub(p, 0xda, 0xe9);
 809 }
 810
 811 void x87_fxch( struct x86_function *p, struct x86_reg arg )
 812 {
 813    assert(arg.file == file_x87);
 814    emit_2ub(p, 0xd9, 0xc8+arg.idx);
 815 }
 816
 817 void x87_fabs( struct x86_function *p )
 818 {
 819    emit_2ub(p, 0xd9, 0xe1);
 820 }
 821
 822 void x87_fchs( struct x86_function *p )
 823 {
 824    emit_2ub(p, 0xd9, 0xe0);
 825 }
 826
 827 void x87_fcos( struct x86_function *p )
 828 {
 829    emit_2ub(p, 0xd9, 0xff);
 830 }
 831
 832
 833 void x87_fprndint( struct x86_function *p )
 834 {
 835    emit_2ub(p, 0xd9, 0xfc);
 836 }
 837
 838 void x87_fscale( struct x86_function *p )
 839 {
 840    emit_2ub(p, 0xd9, 0xfd);
 841 }
 842
 843 void x87_fsin( struct x86_function *p )
 844 {
 845    emit_2ub(p, 0xd9, 0xfe);
 846 }
 847
 848 void x87_fsincos( struct x86_function *p )
 849 {
 850    emit_2ub(p, 0xd9, 0xfb);
 851 }
 852
 853 void x87_fsqrt( struct x86_function *p )
 854 {
 855    emit_2ub(p, 0xd9, 0xfa);
 856 }
 857
 858 void x87_fxtract( struct x86_function *p )
 859 {
 860    emit_2ub(p, 0xd9, 0xf4);
 861 }
 862
 863 /* st0 = (2^st0)-1
 864  *
 865  * Restrictions: -1.0 <= st0 <= 1.0
 866  */
 867 void x87_f2xm1( struct x86_function *p )
 868 {
 869    emit_2ub(p, 0xd9, 0xf0);
 870 }
 871
 872 /* st1 = st1 * log2(st0);
 873  * pop_stack;
 874  */
 875 void x87_fyl2x( struct x86_function *p )
 876 {
 877    emit_2ub(p, 0xd9, 0xf1);
 878 }
 879
 880 /* st1 = st1 * log2(st0 + 1.0);
 881  * pop_stack;
 882  *
 883  * A fast operation, with restrictions: -.29 < st0 < .29
 884  */
 885 void x87_fyl2xp1( struct x86_function *p )
 886 {
 887    emit_2ub(p, 0xd9, 0xf9);
 888 }
 889
 890
 891 void x87_fld( struct x86_function *p, struct x86_reg arg )
 892 {
 893    if (arg.file == file_x87)
 894       emit_2ub(p, 0xd9, 0xc0 + arg.idx);
 895    else {
 896       emit_1ub(p, 0xd9);
 897       emit_modrm_noreg(p, 0, arg);
 898    }
 899 }
 900
 901 void x87_fst( struct x86_function *p, struct x86_reg dst )
 902 {
 903    if (dst.file == file_x87)
 904       emit_2ub(p, 0xdd, 0xd0 + dst.idx);
 905    else {
 906       emit_1ub(p, 0xd9);
 907       emit_modrm_noreg(p, 2, dst);
 908    }
 909 }
 910
 911 void x87_fstp( struct x86_function *p, struct x86_reg dst )
 912 {
 913    if (dst.file == file_x87)
 914       emit_2ub(p, 0xdd, 0xd8 + dst.idx);
 915    else {
 916       emit_1ub(p, 0xd9);
 917       emit_modrm_noreg(p, 3, dst);
 918    }
 919 }
 920
 921 void x87_fcom( struct x86_function *p, struct x86_reg dst )
 922 {
 923    if (dst.file == file_x87)
 924       emit_2ub(p, 0xd8, 0xd0 + dst.idx);
 925    else {
 926       emit_1ub(p, 0xd8);
 927       emit_modrm_noreg(p, 2, dst);
 928    }
 929 }
 930
 931 void x87_fcomp( struct x86_function *p, struct x86_reg dst )
 932 {
 933    if (dst.file == file_x87)
 934       emit_2ub(p, 0xd8, 0xd8 + dst.idx);
 935    else {
 936       emit_1ub(p, 0xd8);
 937       emit_modrm_noreg(p, 3, dst);
 938    }
 939 }
 940
 941
 942 void x87_fnstsw( struct x86_function *p, struct x86_reg dst )
 943 {
 944    assert(dst.file == file_REG32);
 945
 946    if (dst.idx == reg_AX &&
 947        dst.mod == mod_REG)
 948       emit_2ub(p, 0xdf, 0xe0);
 949    else {
 950       emit_1ub(p, 0xdd);
 951       emit_modrm_noreg(p, 7, dst);
 952    }
 953 }
 954
 955
 956
 957
 958 /***********************************************************************
 959  * MMX instructions
 960  */
 961
 962 void mmx_emms( struct x86_function *p )
 963 {
 964    assert(p->need_emms);
 965    emit_2ub(p, 0x0f, 0x77);
 966    p->need_emms = 0;
 967 }
 968
 969 void mmx_packssdw( struct x86_function *p,
 970                    struct x86_reg dst,
 971                    struct x86_reg src )
 972 {
 973    assert(dst.file == file_MMX &&
 974           (src.file == file_MMX || src.mod != mod_REG));
 975
 976    p->need_emms = 1;
 977
 978    emit_2ub(p, X86_TWOB, 0x6b);
 979    emit_modrm( p, dst, src );
 980 }
 981
 982 void mmx_packuswb( struct x86_function *p,
 983                    struct x86_reg dst,
 984                    struct x86_reg src )
 985 {
 986    assert(dst.file == file_MMX &&
 987           (src.file == file_MMX || src.mod != mod_REG));
 988
 989    p->need_emms = 1;
 990
 991    emit_2ub(p, X86_TWOB, 0x67);
 992    emit_modrm( p, dst, src );
 993 }
 994
 995 void mmx_movd( struct x86_function *p,
 996                struct x86_reg dst,
 997                struct x86_reg src )
 998 {
 999    p->need_emms = 1;
1000    emit_1ub(p, X86_TWOB);
1001    emit_op_modrm( p, 0x6e, 0x7e, dst, src );
1002 }
1003
1004 void mmx_movq( struct x86_function *p,
1005                struct x86_reg dst,
1006                struct x86_reg src )
1007 {
1008    p->need_emms = 1;
1009    emit_1ub(p, X86_TWOB);
1010    emit_op_modrm( p, 0x6f, 0x7f, dst, src );
1011 }
1012
1013
1014 /***********************************************************************
1015  * Helper functions
1016  */
1017
1018
1019 /* Retreive a reference to one of the function arguments, taking into
1020  * account any push/pop activity:
1021  */
1022 struct x86_reg x86_fn_arg( struct x86_function *p,
1023                            GLuint arg )
1024 {
1025    return x86_make_disp(x86_make_reg(file_REG32, reg_SP),
1026                         p->stack_offset + arg * 4);     /* ??? */
1027 }
1028
1029
1030 void x86_init_func( struct x86_function *p )
1031 {
1032    x86_init_func_size(p, 1024);
1033 }
1034
1035 void x86_init_func_size( struct x86_function *p, GLuint code_size )
1036 {
1037    p->store = _mesa_exec_malloc(code_size);
1038    p->csr = p->store;
1039 }
1040
1041 void x86_release_func( struct x86_function *p )
1042 {
1043    _mesa_exec_free(p->store);
1044 }
1045
1046
1047 void (*x86_get_func( struct x86_function *p ))(void)
1048 {
1049    if (DISASSEM)
1050       _mesa_printf("disassemble %p %p\n", p->store, p->csr);
1051    return (void (*)(void))p->store;
1052 }
1053
1054 #else
1055
1056 void x86sse_dummy( void )
1057 {
1058 }
1059
1060 #endif