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 void x86_or( struct x86_function *p,
 371              struct x86_reg dst,
 372              struct x86_reg src )
 373 {
 374    emit_op_modrm( p, 0x0b, 0x09, dst, src );
 375 }
 376
 377 void x86_and( struct x86_function *p,
 378               struct x86_reg dst,
 379               struct x86_reg src )
 380 {
 381    emit_op_modrm( p, 0x23, 0x21, dst, src );
 382 }
 383
 384
 385
 386 /***********************************************************************
 387  * SSE instructions
 388  */
 389
 390
 391 void sse_movss( struct x86_function *p,
 392                 struct x86_reg dst,
 393                 struct x86_reg src )
 394 {
 395    emit_2ub(p, 0xF3, X86_TWOB);
 396    emit_op_modrm( p, 0x10, 0x11, dst, src );
 397 }
 398
 399 void sse_movaps( struct x86_function *p,
 400                  struct x86_reg dst,
 401                  struct x86_reg src )
 402 {
 403    emit_1ub(p, X86_TWOB);
 404    emit_op_modrm( p, 0x28, 0x29, dst, src );
 405 }
 406
 407 void sse_movups( struct x86_function *p,
 408                  struct x86_reg dst,
 409                  struct x86_reg src )
 410 {
 411    emit_1ub(p, X86_TWOB);
 412    emit_op_modrm( p, 0x10, 0x11, dst, src );
 413 }
 414
 415 void sse_movhps( struct x86_function *p,
 416                  struct x86_reg dst,
 417                  struct x86_reg src )
 418 {
 419    assert(dst.mod != mod_REG || src.mod != mod_REG);
 420    emit_1ub(p, X86_TWOB);
 421    emit_op_modrm( p, 0x16, 0x17, dst, src ); /* cf movlhps */
 422 }
 423
 424 void sse_movlps( struct x86_function *p,
 425                  struct x86_reg dst,
 426                  struct x86_reg src )
 427 {
 428    assert(dst.mod != mod_REG || src.mod != mod_REG);
 429    emit_1ub(p, X86_TWOB);
 430    emit_op_modrm( p, 0x12, 0x13, dst, src ); /* cf movhlps */
 431 }
 432
 433 void sse_maxps( struct x86_function *p,
 434                 struct x86_reg dst,
 435                 struct x86_reg src )
 436 {
 437    emit_2ub(p, X86_TWOB, 0x5F);
 438    emit_modrm( p, dst, src );
 439 }
 440
 441 void sse_maxss( struct x86_function *p,
 442                 struct x86_reg dst,
 443                 struct x86_reg src )
 444 {
 445    emit_3ub(p, 0xF3, X86_TWOB, 0x5F);
 446    emit_modrm( p, dst, src );
 447 }
 448
 449 void sse_divss( struct x86_function *p,
 450                 struct x86_reg dst,
 451                 struct x86_reg src )
 452 {
 453    emit_3ub(p, 0xF3, X86_TWOB, 0x5E);
 454    emit_modrm( p, dst, src );
 455 }
 456
 457 void sse_minps( struct x86_function *p,
 458                 struct x86_reg dst,
 459                 struct x86_reg src )
 460 {
 461    emit_2ub(p, X86_TWOB, 0x5D);
 462    emit_modrm( p, dst, src );
 463 }
 464
 465 void sse_subps( struct x86_function *p,
 466                 struct x86_reg dst,
 467                 struct x86_reg src )
 468 {
 469    emit_2ub(p, X86_TWOB, 0x5C);
 470    emit_modrm( p, dst, src );
 471 }
 472
 473 void sse_mulps( struct x86_function *p,
 474                 struct x86_reg dst,
 475                 struct x86_reg src )
 476 {
 477    emit_2ub(p, X86_TWOB, 0x59);
 478    emit_modrm( p, dst, src );
 479 }
 480
 481 void sse_mulss( struct x86_function *p,
 482                 struct x86_reg dst,
 483                 struct x86_reg src )
 484 {
 485    emit_3ub(p, 0xF3, X86_TWOB, 0x59);
 486    emit_modrm( p, dst, src );
 487 }
 488
 489 void sse_addps( struct x86_function *p,
 490                 struct x86_reg dst,
 491                 struct x86_reg src )
 492 {
 493    emit_2ub(p, X86_TWOB, 0x58);
 494    emit_modrm( p, dst, src );
 495 }
 496
 497 void sse_addss( struct x86_function *p,
 498                 struct x86_reg dst,
 499                 struct x86_reg src )
 500 {
 501    emit_3ub(p, 0xF3, X86_TWOB, 0x58);
 502    emit_modrm( p, dst, src );
 503 }
 504
 505 void sse_andps( struct x86_function *p,
 506                 struct x86_reg dst,
 507                 struct x86_reg src )
 508 {
 509    emit_2ub(p, X86_TWOB, 0x54);
 510    emit_modrm( p, dst, src );
 511 }
 512
 513
 514 void sse_rsqrtss( struct x86_function *p,
 515                   struct x86_reg dst,
 516                   struct x86_reg src )
 517 {
 518    emit_3ub(p, 0xF3, X86_TWOB, 0x52);
 519    emit_modrm( p, dst, src );
 520
 521 }
 522
 523 void sse_movhlps( struct x86_function *p,
 524                   struct x86_reg dst,
 525                   struct x86_reg src )
 526 {
 527    assert(dst.mod == mod_REG && src.mod == mod_REG);
 528    emit_2ub(p, X86_TWOB, 0x12);
 529    emit_modrm( p, dst, src );
 530 }
 531
 532 void sse_movlhps( struct x86_function *p,
 533                   struct x86_reg dst,
 534                   struct x86_reg src )
 535 {
 536    assert(dst.mod == mod_REG && src.mod == mod_REG);
 537    emit_2ub(p, X86_TWOB, 0x16);
 538    emit_modrm( p, dst, src );
 539 }
 540
 541
 542 void sse_cvtps2pi( struct x86_function *p,
 543                    struct x86_reg dst,
 544                    struct x86_reg src )
 545 {
 546    assert(dst.file == file_MMX &&
 547           (src.file == file_XMM || src.mod != mod_REG));
 548
 549    p->need_emms = 1;
 550
 551    emit_2ub(p, X86_TWOB, 0x2d);
 552    emit_modrm( p, dst, src );
 553 }
 554
 555
 556 /* Shufps can also be used to implement a reduced swizzle when dest ==
 557  * arg0.
 558  */
 559 void sse_shufps( struct x86_function *p,
 560                  struct x86_reg dest,
 561                  struct x86_reg arg0,
 562                  GLubyte shuf)
 563 {
 564    emit_2ub(p, X86_TWOB, 0xC6);
 565    emit_modrm(p, dest, arg0);
 566    emit_1ub(p, shuf);
 567 }
 568
 569 void sse_cmpps( struct x86_function *p,
 570                 struct x86_reg dest,
 571                 struct x86_reg arg0,
 572                 GLubyte cc)
 573 {
 574    emit_2ub(p, X86_TWOB, 0xC2);
 575    emit_modrm(p, dest, arg0);
 576    emit_1ub(p, cc);
 577 }
 578
 579 /***********************************************************************
 580  * SSE2 instructions
 581  */
 582
 583 /**
 584  * Perform a reduced swizzle:
 585  */
 586 void sse2_pshufd( struct x86_function *p,
 587                   struct x86_reg dest,
 588                   struct x86_reg arg0,
 589                   GLubyte shuf)
 590 {
 591    emit_3ub(p, 0x66, X86_TWOB, 0x70);
 592    emit_modrm(p, dest, arg0);
 593    emit_1ub(p, shuf);
 594 }
 595
 596 void sse2_cvtps2dq( struct x86_function *p,
 597                     struct x86_reg dst,
 598                     struct x86_reg src )
 599 {
 600    emit_3ub(p, 0x66, X86_TWOB, 0x5B);
 601    emit_modrm( p, dst, src );
 602 }
 603
 604 void sse2_packssdw( struct x86_function *p,
 605                     struct x86_reg dst,
 606                     struct x86_reg src )
 607 {
 608    emit_3ub(p, 0x66, X86_TWOB, 0x6B);
 609    emit_modrm( p, dst, src );
 610 }
 611
 612 void sse2_packsswb( struct x86_function *p,
 613                     struct x86_reg dst,
 614                     struct x86_reg src )
 615 {
 616    emit_3ub(p, 0x66, X86_TWOB, 0x63);
 617    emit_modrm( p, dst, src );
 618 }
 619
 620 void sse2_packuswb( struct x86_function *p,
 621                     struct x86_reg dst,
 622                     struct x86_reg src )
 623 {
 624    emit_3ub(p, 0x66, X86_TWOB, 0x67);
 625    emit_modrm( p, dst, src );
 626 }
 627
 628 void sse2_rcpss( struct x86_function *p,
 629                 struct x86_reg dst,
 630                 struct x86_reg src )
 631 {
 632    emit_3ub(p, 0xF3, X86_TWOB, 0x53);
 633    emit_modrm( p, dst, src );
 634 }
 635
 636 void sse2_movd( struct x86_function *p,
 637                 struct x86_reg dst,
 638                 struct x86_reg src )
 639 {
 640    emit_2ub(p, 0x66, X86_TWOB);
 641    emit_op_modrm( p, 0x6e, 0x7e, dst, src );
 642 }
 643
 644
 645
 646
 647 /***********************************************************************
 648  * x87 instructions
 649  */
 650 void x87_fist( struct x86_function *p, struct x86_reg dst )
 651 {
 652    emit_1ub(p, 0xdb);
 653    emit_modrm_noreg(p, 2, dst);
 654 }
 655
 656 void x87_fistp( struct x86_function *p, struct x86_reg dst )
 657 {
 658    emit_1ub(p, 0xdb);
 659    emit_modrm_noreg(p, 3, dst);
 660 }
 661
 662 void x87_fild( struct x86_function *p, struct x86_reg arg )
 663 {
 664    emit_1ub(p, 0xdf);
 665    emit_modrm_noreg(p, 0, arg);
 666 }
 667
 668 void x87_fldz( struct x86_function *p )
 669 {
 670    emit_2ub(p, 0xd9, 0xee);
 671 }
 672
 673
 674 void x87_fldcw( struct x86_function *p, struct x86_reg arg )
 675 {
 676    assert(arg.file == file_REG32);
 677    assert(arg.mod != mod_REG);
 678    emit_1ub(p, 0xd9);
 679    emit_modrm_noreg(p, 5, arg);
 680 }
 681
 682 void x87_fld1( struct x86_function *p )
 683 {
 684    emit_2ub(p, 0xd9, 0xe8);
 685 }
 686
 687 void x87_fldl2e( struct x86_function *p )
 688 {
 689    emit_2ub(p, 0xd9, 0xea);
 690 }
 691
 692 void x87_fldln2( struct x86_function *p )
 693 {
 694    emit_2ub(p, 0xd9, 0xed);
 695 }
 696
 697 void x87_fwait( struct x86_function *p )
 698 {
 699    emit_1ub(p, 0x9b);
 700 }
 701
 702 void x87_fnclex( struct x86_function *p )
 703 {
 704    emit_2ub(p, 0xdb, 0xe2);
 705 }
 706
 707 void x87_fclex( struct x86_function *p )
 708 {
 709    x87_fwait(p);
 710    x87_fnclex(p);
 711 }
 712
 713
 714 static void x87_arith_op( struct x86_function *p, struct x86_reg dst, struct x86_reg arg,
 715                           GLubyte dst0ub0,
 716                           GLubyte dst0ub1,
 717                           GLubyte arg0ub0,
 718                           GLubyte arg0ub1,
 719                           GLubyte argmem_noreg)
 720 {
 721    assert(dst.file == file_x87);
 722
 723    if (arg.file == file_x87) {
 724       if (dst.idx == 0)
 725          emit_2ub(p, dst0ub0, dst0ub1+arg.idx);
 726       else if (arg.idx == 0)
 727          emit_2ub(p, arg0ub0, arg0ub1+arg.idx);
 728       else
 729          assert(0);
 730    }
 731    else if (dst.idx == 0) {
 732       assert(arg.file = file_REG32);
 733       emit_1ub(p, 0xd8);
 734       emit_modrm_noreg(p, argmem_noreg, arg);
 735    }
 736    else
 737       assert(0);
 738 }
 739
 740 void x87_fmul( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 741 {
 742    x87_arith_op(p, dst, arg,
 743                 0xd8, 0xc8,
 744                 0xdc, 0xc8,
 745                 4);
 746 }
 747
 748 void x87_fsub( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 749 {
 750    x87_arith_op(p, dst, arg,
 751                 0xd8, 0xe0,
 752                 0xdc, 0xe8,
 753                 4);
 754 }
 755
 756 void x87_fsubr( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 757 {
 758    x87_arith_op(p, dst, arg,
 759                 0xd8, 0xe8,
 760                 0xdc, 0xe0,
 761                 5);
 762 }
 763
 764 void x87_fadd( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 765 {
 766    x87_arith_op(p, dst, arg,
 767                 0xd8, 0xc0,
 768                 0xdc, 0xc0,
 769                 0);
 770 }
 771
 772 void x87_fdiv( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 773 {
 774    x87_arith_op(p, dst, arg,
 775                 0xd8, 0xf0,
 776                 0xdc, 0xf8,
 777                 6);
 778 }
 779
 780 void x87_fdivr( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 781 {
 782    x87_arith_op(p, dst, arg,
 783                 0xd8, 0xf8,
 784                 0xdc, 0xf0,
 785                 7);
 786 }
 787
 788 void x87_fmulp( struct x86_function *p, struct x86_reg dst )
 789 {
 790    assert(dst.file == file_x87);
 791    assert(dst.idx >= 1);
 792    emit_2ub(p, 0xde, 0xc8+dst.idx);
 793 }
 794
 795 void x87_fsubp( struct x86_function *p, struct x86_reg dst )
 796 {
 797    assert(dst.file == file_x87);
 798    assert(dst.idx >= 1);
 799    emit_2ub(p, 0xde, 0xe8+dst.idx);
 800 }
 801
 802 void x87_fsubrp( struct x86_function *p, struct x86_reg dst )
 803 {
 804    assert(dst.file == file_x87);
 805    assert(dst.idx >= 1);
 806    emit_2ub(p, 0xde, 0xe0+dst.idx);
 807 }
 808
 809 void x87_faddp( struct x86_function *p, struct x86_reg dst )
 810 {
 811    assert(dst.file == file_x87);
 812    assert(dst.idx >= 1);
 813    emit_2ub(p, 0xde, 0xc0+dst.idx);
 814 }
 815
 816 void x87_fdivp( struct x86_function *p, struct x86_reg dst )
 817 {
 818    assert(dst.file == file_x87);
 819    assert(dst.idx >= 1);
 820    emit_2ub(p, 0xde, 0xf8+dst.idx);
 821 }
 822
 823 void x87_fdivrp( struct x86_function *p, struct x86_reg dst )
 824 {
 825    assert(dst.file == file_x87);
 826    assert(dst.idx >= 1);
 827    emit_2ub(p, 0xde, 0xf0+dst.idx);
 828 }
 829
 830 void x87_fucom( struct x86_function *p, struct x86_reg arg )
 831 {
 832    assert(arg.file == file_x87);
 833    emit_2ub(p, 0xdd, 0xe0+arg.idx);
 834 }
 835
 836 void x87_fucomp( struct x86_function *p, struct x86_reg arg )
 837 {
 838    assert(arg.file == file_x87);
 839    emit_2ub(p, 0xdd, 0xe8+arg.idx);
 840 }
 841
 842 void x87_fucompp( struct x86_function *p )
 843 {
 844    emit_2ub(p, 0xda, 0xe9);
 845 }
 846
 847 void x87_fxch( struct x86_function *p, struct x86_reg arg )
 848 {
 849    assert(arg.file == file_x87);
 850    emit_2ub(p, 0xd9, 0xc8+arg.idx);
 851 }
 852
 853 void x87_fabs( struct x86_function *p )
 854 {
 855    emit_2ub(p, 0xd9, 0xe1);
 856 }
 857
 858 void x87_fchs( struct x86_function *p )
 859 {
 860    emit_2ub(p, 0xd9, 0xe0);
 861 }
 862
 863 void x87_fcos( struct x86_function *p )
 864 {
 865    emit_2ub(p, 0xd9, 0xff);
 866 }
 867
 868
 869 void x87_fprndint( struct x86_function *p )
 870 {
 871    emit_2ub(p, 0xd9, 0xfc);
 872 }
 873
 874 void x87_fscale( struct x86_function *p )
 875 {
 876    emit_2ub(p, 0xd9, 0xfd);
 877 }
 878
 879 void x87_fsin( struct x86_function *p )
 880 {
 881    emit_2ub(p, 0xd9, 0xfe);
 882 }
 883
 884 void x87_fsincos( struct x86_function *p )
 885 {
 886    emit_2ub(p, 0xd9, 0xfb);
 887 }
 888
 889 void x87_fsqrt( struct x86_function *p )
 890 {
 891    emit_2ub(p, 0xd9, 0xfa);
 892 }
 893
 894 void x87_fxtract( struct x86_function *p )
 895 {
 896    emit_2ub(p, 0xd9, 0xf4);
 897 }
 898
 899 /* st0 = (2^st0)-1
 900  *
 901  * Restrictions: -1.0 <= st0 <= 1.0
 902  */
 903 void x87_f2xm1( struct x86_function *p )
 904 {
 905    emit_2ub(p, 0xd9, 0xf0);
 906 }
 907
 908 /* st1 = st1 * log2(st0);
 909  * pop_stack;
 910  */
 911 void x87_fyl2x( struct x86_function *p )
 912 {
 913    emit_2ub(p, 0xd9, 0xf1);
 914 }
 915
 916 /* st1 = st1 * log2(st0 + 1.0);
 917  * pop_stack;
 918  *
 919  * A fast operation, with restrictions: -.29 < st0 < .29
 920  */
 921 void x87_fyl2xp1( struct x86_function *p )
 922 {
 923    emit_2ub(p, 0xd9, 0xf9);
 924 }
 925
 926
 927 void x87_fld( struct x86_function *p, struct x86_reg arg )
 928 {
 929    if (arg.file == file_x87)
 930       emit_2ub(p, 0xd9, 0xc0 + arg.idx);
 931    else {
 932       emit_1ub(p, 0xd9);
 933       emit_modrm_noreg(p, 0, arg);
 934    }
 935 }
 936
 937 void x87_fst( struct x86_function *p, struct x86_reg dst )
 938 {
 939    if (dst.file == file_x87)
 940       emit_2ub(p, 0xdd, 0xd0 + dst.idx);
 941    else {
 942       emit_1ub(p, 0xd9);
 943       emit_modrm_noreg(p, 2, dst);
 944    }
 945 }
 946
 947 void x87_fstp( struct x86_function *p, struct x86_reg dst )
 948 {
 949    if (dst.file == file_x87)
 950       emit_2ub(p, 0xdd, 0xd8 + dst.idx);
 951    else {
 952       emit_1ub(p, 0xd9);
 953       emit_modrm_noreg(p, 3, dst);
 954    }
 955 }
 956
 957 void x87_fcom( struct x86_function *p, struct x86_reg dst )
 958 {
 959    if (dst.file == file_x87)
 960       emit_2ub(p, 0xd8, 0xd0 + dst.idx);
 961    else {
 962       emit_1ub(p, 0xd8);
 963       emit_modrm_noreg(p, 2, dst);
 964    }
 965 }
 966
 967 void x87_fcomp( struct x86_function *p, struct x86_reg dst )
 968 {
 969    if (dst.file == file_x87)
 970       emit_2ub(p, 0xd8, 0xd8 + dst.idx);
 971    else {
 972       emit_1ub(p, 0xd8);
 973       emit_modrm_noreg(p, 3, dst);
 974    }
 975 }
 976
 977
 978 void x87_fnstsw( struct x86_function *p, struct x86_reg dst )
 979 {
 980    assert(dst.file == file_REG32);
 981
 982    if (dst.idx == reg_AX &&
 983        dst.mod == mod_REG)
 984       emit_2ub(p, 0xdf, 0xe0);
 985    else {
 986       emit_1ub(p, 0xdd);
 987       emit_modrm_noreg(p, 7, dst);
 988    }
 989 }
 990
 991
 992
 993
 994 /***********************************************************************
 995  * MMX instructions
 996  */
 997
 998 void mmx_emms( struct x86_function *p )
 999 {
1000    assert(p->need_emms);
1001    emit_2ub(p, 0x0f, 0x77);
1002    p->need_emms = 0;
1003 }
1004
1005 void mmx_packssdw( struct x86_function *p,
1006                    struct x86_reg dst,
1007                    struct x86_reg src )
1008 {
1009    assert(dst.file == file_MMX &&
1010           (src.file == file_MMX || src.mod != mod_REG));
1011
1012    p->need_emms = 1;
1013
1014    emit_2ub(p, X86_TWOB, 0x6b);
1015    emit_modrm( p, dst, src );
1016 }
1017
1018 void mmx_packuswb( struct x86_function *p,
1019                    struct x86_reg dst,
1020                    struct x86_reg src )
1021 {
1022    assert(dst.file == file_MMX &&
1023           (src.file == file_MMX || src.mod != mod_REG));
1024
1025    p->need_emms = 1;
1026
1027    emit_2ub(p, X86_TWOB, 0x67);
1028    emit_modrm( p, dst, src );
1029 }
1030
1031 void mmx_movd( struct x86_function *p,
1032                struct x86_reg dst,
1033                struct x86_reg src )
1034 {
1035    p->need_emms = 1;
1036    emit_1ub(p, X86_TWOB);
1037    emit_op_modrm( p, 0x6e, 0x7e, dst, src );
1038 }
1039
1040 void mmx_movq( struct x86_function *p,
1041                struct x86_reg dst,
1042                struct x86_reg src )
1043 {
1044    p->need_emms = 1;
1045    emit_1ub(p, X86_TWOB);
1046    emit_op_modrm( p, 0x6f, 0x7f, dst, src );
1047 }
1048
1049
1050 /***********************************************************************
1051  * Helper functions
1052  */
1053
1054
1055 /* Retreive a reference to one of the function arguments, taking into
1056  * account any push/pop activity:
1057  */
1058 struct x86_reg x86_fn_arg( struct x86_function *p,
1059                            GLuint arg )
1060 {
1061    return x86_make_disp(x86_make_reg(file_REG32, reg_SP),
1062                         p->stack_offset + arg * 4);     /* ??? */
1063 }
1064
1065
1066 void x86_init_func( struct x86_function *p )
1067 {
1068    x86_init_func_size(p, 1024);
1069 }
1070
1071 void x86_init_func_size( struct x86_function *p, GLuint code_size )
1072 {
1073    p->store = _mesa_exec_malloc(code_size);
1074    p->csr = p->store;
1075 }
1076
1077 void x86_release_func( struct x86_function *p )
1078 {
1079    _mesa_exec_free(p->store);
1080 }
1081
1082
1083 void (*x86_get_func( struct x86_function *p ))(void)
1084 {
1085    if (DISASSEM)
1086       _mesa_printf("disassemble %p %p\n", p->store, p->csr);
1087    return (void (*)(void))p->store;
1088 }
1089
1090 #else
1091
1092 void x86sse_dummy( void )
1093 {
1094 }
1095
1096 #endif