src/mesa/x86/rtasm/x86sse.c

   1 #if defined(__i386__) || defined(__386__)
   2
   3 #include "imports.h"
   4 #include "x86sse.h"
   5
   6 #define DISASSEM 0
   7 #define X86_TWOB 0x0f
   8
   9 static GLubyte *cptr( void (*label)() )
  10 {
  11    return (char *)(unsigned long)label;
  12 }
  13
  14
  15 /* Emit bytes to the instruction stream:
  16  */
  17 static void emit_1b( struct x86_function *p, GLbyte b0 )
  18 {
  19    *(GLbyte *)(p->csr++) = b0;
  20 }
  21
  22 static void emit_1i( struct x86_function *p, GLint i0 )
  23 {
  24    *(GLint *)(p->csr) = i0;
  25    p->csr += 4;
  26 }
  27
  28 static void disassem( struct x86_function *p, const char *fn )
  29 {
  30 #if DISASSEM && 0
  31    if (fn && fn != p->fn) {
  32       _mesa_printf("0x%x: %s\n", p->csr, fn);
  33       p->fn = fn;
  34    }
  35 #endif
  36 }
  37
  38 static void emit_1ub_fn( struct x86_function *p, GLubyte b0, const char *fn )
  39 {
  40    disassem(p, fn);
  41    *(p->csr++) = b0;
  42 }
  43
  44 static void emit_2ub_fn( struct x86_function *p, GLubyte b0, GLubyte b1, const char *fn )
  45 {
  46    disassem(p, fn);
  47    *(p->csr++) = b0;
  48    *(p->csr++) = b1;
  49 }
  50
  51 static void emit_3ub_fn( struct x86_function *p, GLubyte b0, GLubyte b1, GLubyte b2, const char *fn )
  52 {
  53    disassem(p, fn);
  54    *(p->csr++) = b0;
  55    *(p->csr++) = b1;
  56    *(p->csr++) = b2;
  57 }
  58
  59 #define emit_1ub(p, b0)         emit_1ub_fn(p, b0, __FUNCTION__)
  60 #define emit_2ub(p, b0, b1)     emit_2ub_fn(p, b0, b1, __FUNCTION__)
  61 #define emit_3ub(p, b0, b1, b2) emit_3ub_fn(p, b0, b1, b2, __FUNCTION__)
  62
  63
  64
  65 /* Build a modRM byte + possible displacement.  No treatment of SIB
  66  * indexing.  BZZT - no way to encode an absolute address.
  67  */
  68 static void emit_modrm( struct x86_function *p,
  69                         struct x86_reg reg,
  70                         struct x86_reg regmem )
  71 {
  72    GLubyte val = 0;
  73
  74    assert(reg.mod == mod_REG);
  75
  76    val |= regmem.mod << 6;      /* mod field */
  77    val |= reg.idx << 3;         /* reg field */
  78    val |= regmem.idx;           /* r/m field */
  79
  80    emit_1ub_fn(p, val, 0);
  81
  82    /* Oh-oh we've stumbled into the SIB thing.
  83     */
  84    if (regmem.file == file_REG32 &&
  85        regmem.idx == reg_SP) {
  86       emit_1ub_fn(p, 0x24, 0);          /* simplistic! */
  87    }
  88
  89    switch (regmem.mod) {
  90    case mod_REG:
  91    case mod_INDIRECT:
  92       break;
  93    case mod_DISP8:
  94       emit_1b(p, regmem.disp);
  95       break;
  96    case mod_DISP32:
  97       emit_1i(p, regmem.disp);
  98       break;
  99    default:
 100       assert(0);
 101       break;
 102    }
 103 }
 104
 105
 106 static void emit_modrm_noreg( struct x86_function *p,
 107                               GLuint op,
 108                               struct x86_reg regmem )
 109 {
 110    struct x86_reg dummy = x86_make_reg(file_REG32, op);
 111    emit_modrm(p, dummy, regmem);
 112 }
 113
 114 /* Many x86 instructions have two opcodes to cope with the situations
 115  * where the destination is a register or memory reference
 116  * respectively.  This function selects the correct opcode based on
 117  * the arguments presented.
 118  */
 119 static void emit_op_modrm( struct x86_function *p,
 120                            GLubyte op_dst_is_reg,
 121                            GLubyte op_dst_is_mem,
 122                            struct x86_reg dst,
 123                            struct x86_reg src )
 124 {
 125    switch (dst.mod) {
 126    case mod_REG:
 127       emit_1ub_fn(p, op_dst_is_reg, 0);
 128       emit_modrm(p, dst, src);
 129       break;
 130    case mod_INDIRECT:
 131    case mod_DISP32:
 132    case mod_DISP8:
 133       assert(src.mod == mod_REG);
 134       emit_1ub_fn(p, op_dst_is_mem, 0);
 135       emit_modrm(p, src, dst);
 136       break;
 137    default:
 138       assert(0);
 139       break;
 140    }
 141 }
 142
 143
 144
 145
 146
 147
 148
 149 /* Create and manipulate registers and regmem values:
 150  */
 151 struct x86_reg x86_make_reg( enum x86_reg_file file,
 152                              enum x86_reg_name idx )
 153 {
 154    struct x86_reg reg;
 155
 156    reg.file = file;
 157    reg.idx = idx;
 158    reg.mod = mod_REG;
 159    reg.disp = 0;
 160
 161    return reg;
 162 }
 163
 164 struct x86_reg x86_make_disp( struct x86_reg reg,
 165                               GLint disp )
 166 {
 167    assert(reg.file == file_REG32);
 168
 169    if (reg.mod == mod_REG)
 170       reg.disp = disp;
 171    else
 172       reg.disp += disp;
 173
 174    if (reg.disp == 0)
 175       reg.mod = mod_INDIRECT;
 176    else if (reg.disp <= 127 && reg.disp >= -128)
 177       reg.mod = mod_DISP8;
 178    else
 179       reg.mod = mod_DISP32;
 180
 181    return reg;
 182 }
 183
 184 struct x86_reg x86_deref( struct x86_reg reg )
 185 {
 186    return x86_make_disp(reg, 0);
 187 }
 188
 189 struct x86_reg x86_get_base_reg( struct x86_reg reg )
 190 {
 191    return x86_make_reg( reg.file, reg.idx );
 192 }
 193
 194 GLubyte *x86_get_label( struct x86_function *p )
 195 {
 196    return p->csr;
 197 }
 198
 199
 200
 201 /***********************************************************************
 202  * x86 instructions
 203  */
 204
 205
 206 void x86_jcc( struct x86_function *p,
 207               enum x86_cc cc,
 208               GLubyte *label )
 209 {
 210    GLint offset = label - (x86_get_label(p) + 2);
 211
 212    if (offset <= 127 && offset >= -128) {
 213       emit_1ub(p, 0x70 + cc);
 214       emit_1b(p, (GLbyte) offset);
 215    }
 216    else {
 217       offset = label - (x86_get_label(p) + 6);
 218       emit_2ub(p, 0x0f, 0x80 + cc);
 219       emit_1i(p, offset);
 220    }
 221 }
 222
 223 /* Always use a 32bit offset for forward jumps:
 224  */
 225 GLubyte *x86_jcc_forward( struct x86_function *p,
 226                           enum x86_cc cc )
 227 {
 228    emit_2ub(p, 0x0f, 0x80 + cc);
 229    emit_1i(p, 0);
 230    return x86_get_label(p);
 231 }
 232
 233 GLubyte *x86_jmp_forward( struct x86_function *p)
 234 {
 235    emit_1ub(p, 0xe9);
 236    emit_1i(p, 0);
 237    return x86_get_label(p);
 238 }
 239
 240 GLubyte *x86_call_forward( struct x86_function *p)
 241 {
 242    emit_1ub(p, 0xe8);
 243    emit_1i(p, 0);
 244    return x86_get_label(p);
 245 }
 246
 247 /* Fixup offset from forward jump:
 248  */
 249 void x86_fixup_fwd_jump( struct x86_function *p,
 250                          GLubyte *fixup )
 251 {
 252    *(int *)(fixup - 4) = x86_get_label(p) - fixup;
 253 }
 254
 255 void x86_jmp( struct x86_function *p, GLubyte *label)
 256 {
 257    emit_1ub(p, 0xe9);
 258    emit_1i(p, label - x86_get_label(p) - 4);
 259 }
 260
 261 void x86_call( struct x86_function *p, void (*label)())
 262 {
 263    emit_1ub(p, 0xe8);
 264    emit_1i(p, cptr(label) - x86_get_label(p) - 4);
 265 }
 266
 267 /* michal:
 268  * Temporary. As I need immediate operands, and dont want to mess with the codegen,
 269  * I load the immediate into general purpose register and use it.
 270  */
 271 void x86_mov_reg_imm( struct x86_function *p, struct x86_reg dst, GLint imm )
 272 {
 273    assert(dst.mod == mod_REG);
 274    emit_1ub(p, 0xb8 + dst.idx);
 275    emit_1i(p, imm);
 276 }
 277
 278 void x86_push( struct x86_function *p,
 279                struct x86_reg reg )
 280 {
 281    assert(reg.mod == mod_REG);
 282    emit_1ub(p, 0x50 + reg.idx);
 283    p->stack_offset += 4;
 284 }
 285
 286 void x86_pop( struct x86_function *p,
 287               struct x86_reg reg )
 288 {
 289    assert(reg.mod == mod_REG);
 290    emit_1ub(p, 0x58 + reg.idx);
 291    p->stack_offset -= 4;
 292 }
 293
 294 void x86_inc( struct x86_function *p,
 295               struct x86_reg reg )
 296 {
 297    assert(reg.mod == mod_REG);
 298    emit_1ub(p, 0x40 + reg.idx);
 299 }
 300
 301 void x86_dec( struct x86_function *p,
 302               struct x86_reg reg )
 303 {
 304    assert(reg.mod == mod_REG);
 305    emit_1ub(p, 0x48 + reg.idx);
 306 }
 307
 308 void x86_ret( struct x86_function *p )
 309 {
 310    emit_1ub(p, 0xc3);
 311 }
 312
 313 void x86_sahf( struct x86_function *p )
 314 {
 315    emit_1ub(p, 0x9e);
 316 }
 317
 318 void x86_mov( struct x86_function *p,
 319               struct x86_reg dst,
 320               struct x86_reg src )
 321 {
 322    emit_op_modrm( p, 0x8b, 0x89, dst, src );
 323 }
 324
 325 void x86_xor( struct x86_function *p,
 326               struct x86_reg dst,
 327               struct x86_reg src )
 328 {
 329    emit_op_modrm( p, 0x33, 0x31, dst, src );
 330 }
 331
 332 void x86_cmp( struct x86_function *p,
 333               struct x86_reg dst,
 334               struct x86_reg src )
 335 {
 336    emit_op_modrm( p, 0x3b, 0x39, dst, src );
 337 }
 338
 339 void x86_lea( struct x86_function *p,
 340               struct x86_reg dst,
 341               struct x86_reg src )
 342 {
 343    emit_1ub(p, 0x8d);
 344    emit_modrm( p, dst, src );
 345 }
 346
 347 void x86_test( struct x86_function *p,
 348                struct x86_reg dst,
 349                struct x86_reg src )
 350 {
 351    emit_1ub(p, 0x85);
 352    emit_modrm( p, dst, src );
 353 }
 354
 355 void x86_add( struct x86_function *p,
 356                struct x86_reg dst,
 357                struct x86_reg src )
 358 {
 359    emit_op_modrm(p, 0x03, 0x01, dst, src );
 360 }
 361
 362 void x86_mul( struct x86_function *p,
 363                struct x86_reg src )
 364 {
 365    assert (src.file == file_REG32 && src.mod == mod_REG);
 366    emit_op_modrm(p, 0xf7, 0, x86_make_reg (file_REG32, reg_SP), src );
 367 }
 368
 369 void x86_sub( struct x86_function *p,
 370                struct x86_reg dst,
 371                struct x86_reg src )
 372 {
 373    emit_op_modrm(p, 0x2b, 0x29, dst, src );
 374 }
 375
 376 void x86_or( struct x86_function *p,
 377              struct x86_reg dst,
 378              struct x86_reg src )
 379 {
 380    emit_op_modrm( p, 0x0b, 0x09, dst, src );
 381 }
 382
 383 void x86_and( struct x86_function *p,
 384               struct x86_reg dst,
 385               struct x86_reg src )
 386 {
 387    emit_op_modrm( p, 0x23, 0x21, dst, src );
 388 }
 389
 390
 391
 392 /***********************************************************************
 393  * SSE instructions
 394  */
 395
 396
 397 void sse_movss( struct x86_function *p,
 398                 struct x86_reg dst,
 399                 struct x86_reg src )
 400 {
 401    emit_2ub(p, 0xF3, X86_TWOB);
 402    emit_op_modrm( p, 0x10, 0x11, dst, src );
 403 }
 404
 405 void sse_movaps( struct x86_function *p,
 406                  struct x86_reg dst,
 407                  struct x86_reg src )
 408 {
 409    emit_1ub(p, X86_TWOB);
 410    emit_op_modrm( p, 0x28, 0x29, dst, src );
 411 }
 412
 413 void sse_movups( struct x86_function *p,
 414                  struct x86_reg dst,
 415                  struct x86_reg src )
 416 {
 417    emit_1ub(p, X86_TWOB);
 418    emit_op_modrm( p, 0x10, 0x11, dst, src );
 419 }
 420
 421 void sse_movhps( struct x86_function *p,
 422                  struct x86_reg dst,
 423                  struct x86_reg src )
 424 {
 425    assert(dst.mod != mod_REG || src.mod != mod_REG);
 426    emit_1ub(p, X86_TWOB);
 427    emit_op_modrm( p, 0x16, 0x17, dst, src ); /* cf movlhps */
 428 }
 429
 430 void sse_movlps( struct x86_function *p,
 431                  struct x86_reg dst,
 432                  struct x86_reg src )
 433 {
 434    assert(dst.mod != mod_REG || src.mod != mod_REG);
 435    emit_1ub(p, X86_TWOB);
 436    emit_op_modrm( p, 0x12, 0x13, dst, src ); /* cf movhlps */
 437 }
 438
 439 void sse_maxps( struct x86_function *p,
 440                 struct x86_reg dst,
 441                 struct x86_reg src )
 442 {
 443    emit_2ub(p, X86_TWOB, 0x5F);
 444    emit_modrm( p, dst, src );
 445 }
 446
 447 void sse_maxss( struct x86_function *p,
 448                 struct x86_reg dst,
 449                 struct x86_reg src )
 450 {
 451    emit_3ub(p, 0xF3, X86_TWOB, 0x5F);
 452    emit_modrm( p, dst, src );
 453 }
 454
 455 void sse_divss( struct x86_function *p,
 456                 struct x86_reg dst,
 457                 struct x86_reg src )
 458 {
 459    emit_3ub(p, 0xF3, X86_TWOB, 0x5E);
 460    emit_modrm( p, dst, src );
 461 }
 462
 463 void sse_minps( struct x86_function *p,
 464                 struct x86_reg dst,
 465                 struct x86_reg src )
 466 {
 467    emit_2ub(p, X86_TWOB, 0x5D);
 468    emit_modrm( p, dst, src );
 469 }
 470
 471 void sse_subps( struct x86_function *p,
 472                 struct x86_reg dst,
 473                 struct x86_reg src )
 474 {
 475    emit_2ub(p, X86_TWOB, 0x5C);
 476    emit_modrm( p, dst, src );
 477 }
 478
 479 void sse_mulps( struct x86_function *p,
 480                 struct x86_reg dst,
 481                 struct x86_reg src )
 482 {
 483    emit_2ub(p, X86_TWOB, 0x59);
 484    emit_modrm( p, dst, src );
 485 }
 486
 487 void sse_mulss( struct x86_function *p,
 488                 struct x86_reg dst,
 489                 struct x86_reg src )
 490 {
 491    emit_3ub(p, 0xF3, X86_TWOB, 0x59);
 492    emit_modrm( p, dst, src );
 493 }
 494
 495 void sse_addps( struct x86_function *p,
 496                 struct x86_reg dst,
 497                 struct x86_reg src )
 498 {
 499    emit_2ub(p, X86_TWOB, 0x58);
 500    emit_modrm( p, dst, src );
 501 }
 502
 503 void sse_addss( struct x86_function *p,
 504                 struct x86_reg dst,
 505                 struct x86_reg src )
 506 {
 507    emit_3ub(p, 0xF3, X86_TWOB, 0x58);
 508    emit_modrm( p, dst, src );
 509 }
 510
 511 void sse_andnps( struct x86_function *p,
 512                  struct x86_reg dst,
 513                  struct x86_reg src )
 514 {
 515    emit_2ub(p, X86_TWOB, 0x55);
 516    emit_modrm( p, dst, src );
 517 }
 518
 519 void sse_andps( struct x86_function *p,
 520                 struct x86_reg dst,
 521                 struct x86_reg src )
 522 {
 523    emit_2ub(p, X86_TWOB, 0x54);
 524    emit_modrm( p, dst, src );
 525 }
 526
 527 void sse_rsqrtps( struct x86_function *p,
 528                   struct x86_reg dst,
 529                   struct x86_reg src )
 530 {
 531    emit_2ub(p, X86_TWOB, 0x52);
 532    emit_modrm( p, dst, src );
 533 }
 534
 535 void sse_rsqrtss( struct x86_function *p,
 536                   struct x86_reg dst,
 537                   struct x86_reg src )
 538 {
 539    emit_3ub(p, 0xF3, X86_TWOB, 0x52);
 540    emit_modrm( p, dst, src );
 541
 542 }
 543
 544 void sse_movhlps( struct x86_function *p,
 545                   struct x86_reg dst,
 546                   struct x86_reg src )
 547 {
 548    assert(dst.mod == mod_REG && src.mod == mod_REG);
 549    emit_2ub(p, X86_TWOB, 0x12);
 550    emit_modrm( p, dst, src );
 551 }
 552
 553 void sse_movlhps( struct x86_function *p,
 554                   struct x86_reg dst,
 555                   struct x86_reg src )
 556 {
 557    assert(dst.mod == mod_REG && src.mod == mod_REG);
 558    emit_2ub(p, X86_TWOB, 0x16);
 559    emit_modrm( p, dst, src );
 560 }
 561
 562 void sse_orps( struct x86_function *p,
 563                struct x86_reg dst,
 564                struct x86_reg src )
 565 {
 566    emit_2ub(p, X86_TWOB, 0x56);
 567    emit_modrm( p, dst, src );
 568 }
 569
 570 void sse_xorps( struct x86_function *p,
 571                 struct x86_reg dst,
 572                 struct x86_reg src )
 573 {
 574    emit_2ub(p, X86_TWOB, 0x57);
 575    emit_modrm( p, dst, src );
 576 }
 577
 578 void sse_cvtps2pi( struct x86_function *p,
 579                    struct x86_reg dst,
 580                    struct x86_reg src )
 581 {
 582    assert(dst.file == file_MMX &&
 583           (src.file == file_XMM || src.mod != mod_REG));
 584
 585    p->need_emms = 1;
 586
 587    emit_2ub(p, X86_TWOB, 0x2d);
 588    emit_modrm( p, dst, src );
 589 }
 590
 591
 592 /* Shufps can also be used to implement a reduced swizzle when dest ==
 593  * arg0.
 594  */
 595 void sse_shufps( struct x86_function *p,
 596                  struct x86_reg dest,
 597                  struct x86_reg arg0,
 598                  GLubyte shuf)
 599 {
 600    emit_2ub(p, X86_TWOB, 0xC6);
 601    emit_modrm(p, dest, arg0);
 602    emit_1ub(p, shuf);
 603 }
 604
 605 void sse_cmpps( struct x86_function *p,
 606                 struct x86_reg dest,
 607                 struct x86_reg arg0,
 608                 GLubyte cc)
 609 {
 610    emit_2ub(p, X86_TWOB, 0xC2);
 611    emit_modrm(p, dest, arg0);
 612    emit_1ub(p, cc);
 613 }
 614
 615 void sse_pmovmskb( struct x86_function *p,
 616                    struct x86_reg dest,
 617                    struct x86_reg src)
 618 {
 619     emit_3ub(p, 0x66, X86_TWOB, 0xD7);
 620     emit_modrm(p, dest, src);
 621 }
 622
 623 /***********************************************************************
 624  * SSE2 instructions
 625  */
 626
 627 /**
 628  * Perform a reduced swizzle:
 629  */
 630 void sse2_pshufd( struct x86_function *p,
 631                   struct x86_reg dest,
 632                   struct x86_reg arg0,
 633                   GLubyte shuf)
 634 {
 635    emit_3ub(p, 0x66, X86_TWOB, 0x70);
 636    emit_modrm(p, dest, arg0);
 637    emit_1ub(p, shuf);
 638 }
 639
 640 void sse2_cvttps2dq( struct x86_function *p,
 641                      struct x86_reg dst,
 642                      struct x86_reg src )
 643 {
 644    emit_3ub( p, 0xF3, X86_TWOB, 0x5B );
 645    emit_modrm( p, dst, src );
 646 }
 647
 648 void sse2_cvtps2dq( struct x86_function *p,
 649                     struct x86_reg dst,
 650                     struct x86_reg src )
 651 {
 652    emit_3ub(p, 0x66, X86_TWOB, 0x5B);
 653    emit_modrm( p, dst, src );
 654 }
 655
 656 void sse2_packssdw( struct x86_function *p,
 657                     struct x86_reg dst,
 658                     struct x86_reg src )
 659 {
 660    emit_3ub(p, 0x66, X86_TWOB, 0x6B);
 661    emit_modrm( p, dst, src );
 662 }
 663
 664 void sse2_packsswb( struct x86_function *p,
 665                     struct x86_reg dst,
 666                     struct x86_reg src )
 667 {
 668    emit_3ub(p, 0x66, X86_TWOB, 0x63);
 669    emit_modrm( p, dst, src );
 670 }
 671
 672 void sse2_packuswb( struct x86_function *p,
 673                     struct x86_reg dst,
 674                     struct x86_reg src )
 675 {
 676    emit_3ub(p, 0x66, X86_TWOB, 0x67);
 677    emit_modrm( p, dst, src );
 678 }
 679
 680 void sse2_rcpps( struct x86_function *p,
 681                  struct x86_reg dst,
 682                  struct x86_reg src )
 683 {
 684    emit_2ub(p, X86_TWOB, 0x53);
 685    emit_modrm( p, dst, src );
 686 }
 687
 688 void sse2_rcpss( struct x86_function *p,
 689                 struct x86_reg dst,
 690                 struct x86_reg src )
 691 {
 692    emit_3ub(p, 0xF3, X86_TWOB, 0x53);
 693    emit_modrm( p, dst, src );
 694 }
 695
 696 void sse2_movd( struct x86_function *p,
 697                 struct x86_reg dst,
 698                 struct x86_reg src )
 699 {
 700    emit_2ub(p, 0x66, X86_TWOB);
 701    emit_op_modrm( p, 0x6e, 0x7e, dst, src );
 702 }
 703
 704
 705
 706
 707 /***********************************************************************
 708  * x87 instructions
 709  */
 710 void x87_fist( struct x86_function *p, struct x86_reg dst )
 711 {
 712    emit_1ub(p, 0xdb);
 713    emit_modrm_noreg(p, 2, dst);
 714 }
 715
 716 void x87_fistp( struct x86_function *p, struct x86_reg dst )
 717 {
 718    emit_1ub(p, 0xdb);
 719    emit_modrm_noreg(p, 3, dst);
 720 }
 721
 722 void x87_fild( struct x86_function *p, struct x86_reg arg )
 723 {
 724    emit_1ub(p, 0xdf);
 725    emit_modrm_noreg(p, 0, arg);
 726 }
 727
 728 void x87_fldz( struct x86_function *p )
 729 {
 730    emit_2ub(p, 0xd9, 0xee);
 731 }
 732
 733
 734 void x87_fldcw( struct x86_function *p, struct x86_reg arg )
 735 {
 736    assert(arg.file == file_REG32);
 737    assert(arg.mod != mod_REG);
 738    emit_1ub(p, 0xd9);
 739    emit_modrm_noreg(p, 5, arg);
 740 }
 741
 742 void x87_fld1( struct x86_function *p )
 743 {
 744    emit_2ub(p, 0xd9, 0xe8);
 745 }
 746
 747 void x87_fldl2e( struct x86_function *p )
 748 {
 749    emit_2ub(p, 0xd9, 0xea);
 750 }
 751
 752 void x87_fldln2( struct x86_function *p )
 753 {
 754    emit_2ub(p, 0xd9, 0xed);
 755 }
 756
 757 void x87_fwait( struct x86_function *p )
 758 {
 759    emit_1ub(p, 0x9b);
 760 }
 761
 762 void x87_fnclex( struct x86_function *p )
 763 {
 764    emit_2ub(p, 0xdb, 0xe2);
 765 }
 766
 767 void x87_fclex( struct x86_function *p )
 768 {
 769    x87_fwait(p);
 770    x87_fnclex(p);
 771 }
 772
 773
 774 static void x87_arith_op( struct x86_function *p, struct x86_reg dst, struct x86_reg arg,
 775                           GLubyte dst0ub0,
 776                           GLubyte dst0ub1,
 777                           GLubyte arg0ub0,
 778                           GLubyte arg0ub1,
 779                           GLubyte argmem_noreg)
 780 {
 781    assert(dst.file == file_x87);
 782
 783    if (arg.file == file_x87) {
 784       if (dst.idx == 0)
 785          emit_2ub(p, dst0ub0, dst0ub1+arg.idx);
 786       else if (arg.idx == 0)
 787          emit_2ub(p, arg0ub0, arg0ub1+arg.idx);
 788       else
 789          assert(0);
 790    }
 791    else if (dst.idx == 0) {
 792       assert(arg.file = file_REG32);
 793       emit_1ub(p, 0xd8);
 794       emit_modrm_noreg(p, argmem_noreg, arg);
 795    }
 796    else
 797       assert(0);
 798 }
 799
 800 void x87_fmul( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 801 {
 802    x87_arith_op(p, dst, arg,
 803                 0xd8, 0xc8,
 804                 0xdc, 0xc8,
 805                 4);
 806 }
 807
 808 void x87_fsub( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 809 {
 810    x87_arith_op(p, dst, arg,
 811                 0xd8, 0xe0,
 812                 0xdc, 0xe8,
 813                 4);
 814 }
 815
 816 void x87_fsubr( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 817 {
 818    x87_arith_op(p, dst, arg,
 819                 0xd8, 0xe8,
 820                 0xdc, 0xe0,
 821                 5);
 822 }
 823
 824 void x87_fadd( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 825 {
 826    x87_arith_op(p, dst, arg,
 827                 0xd8, 0xc0,
 828                 0xdc, 0xc0,
 829                 0);
 830 }
 831
 832 void x87_fdiv( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 833 {
 834    x87_arith_op(p, dst, arg,
 835                 0xd8, 0xf0,
 836                 0xdc, 0xf8,
 837                 6);
 838 }
 839
 840 void x87_fdivr( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 841 {
 842    x87_arith_op(p, dst, arg,
 843                 0xd8, 0xf8,
 844                 0xdc, 0xf0,
 845                 7);
 846 }
 847
 848 void x87_fmulp( struct x86_function *p, struct x86_reg dst )
 849 {
 850    assert(dst.file == file_x87);
 851    assert(dst.idx >= 1);
 852    emit_2ub(p, 0xde, 0xc8+dst.idx);
 853 }
 854
 855 void x87_fsubp( struct x86_function *p, struct x86_reg dst )
 856 {
 857    assert(dst.file == file_x87);
 858    assert(dst.idx >= 1);
 859    emit_2ub(p, 0xde, 0xe8+dst.idx);
 860 }
 861
 862 void x87_fsubrp( struct x86_function *p, struct x86_reg dst )
 863 {
 864    assert(dst.file == file_x87);
 865    assert(dst.idx >= 1);
 866    emit_2ub(p, 0xde, 0xe0+dst.idx);
 867 }
 868
 869 void x87_faddp( struct x86_function *p, struct x86_reg dst )
 870 {
 871    assert(dst.file == file_x87);
 872    assert(dst.idx >= 1);
 873    emit_2ub(p, 0xde, 0xc0+dst.idx);
 874 }
 875
 876 void x87_fdivp( struct x86_function *p, struct x86_reg dst )
 877 {
 878    assert(dst.file == file_x87);
 879    assert(dst.idx >= 1);
 880    emit_2ub(p, 0xde, 0xf8+dst.idx);
 881 }
 882
 883 void x87_fdivrp( struct x86_function *p, struct x86_reg dst )
 884 {
 885    assert(dst.file == file_x87);
 886    assert(dst.idx >= 1);
 887    emit_2ub(p, 0xde, 0xf0+dst.idx);
 888 }
 889
 890 void x87_fucom( struct x86_function *p, struct x86_reg arg )
 891 {
 892    assert(arg.file == file_x87);
 893    emit_2ub(p, 0xdd, 0xe0+arg.idx);
 894 }
 895
 896 void x87_fucomp( struct x86_function *p, struct x86_reg arg )
 897 {
 898    assert(arg.file == file_x87);
 899    emit_2ub(p, 0xdd, 0xe8+arg.idx);
 900 }
 901
 902 void x87_fucompp( struct x86_function *p )
 903 {
 904    emit_2ub(p, 0xda, 0xe9);
 905 }
 906
 907 void x87_fxch( struct x86_function *p, struct x86_reg arg )
 908 {
 909    assert(arg.file == file_x87);
 910    emit_2ub(p, 0xd9, 0xc8+arg.idx);
 911 }
 912
 913 void x87_fabs( struct x86_function *p )
 914 {
 915    emit_2ub(p, 0xd9, 0xe1);
 916 }
 917
 918 void x87_fchs( struct x86_function *p )
 919 {
 920    emit_2ub(p, 0xd9, 0xe0);
 921 }
 922
 923 void x87_fcos( struct x86_function *p )
 924 {
 925    emit_2ub(p, 0xd9, 0xff);
 926 }
 927
 928
 929 void x87_fprndint( struct x86_function *p )
 930 {
 931    emit_2ub(p, 0xd9, 0xfc);
 932 }
 933
 934 void x87_fscale( struct x86_function *p )
 935 {
 936    emit_2ub(p, 0xd9, 0xfd);
 937 }
 938
 939 void x87_fsin( struct x86_function *p )
 940 {
 941    emit_2ub(p, 0xd9, 0xfe);
 942 }
 943
 944 void x87_fsincos( struct x86_function *p )
 945 {
 946    emit_2ub(p, 0xd9, 0xfb);
 947 }
 948
 949 void x87_fsqrt( struct x86_function *p )
 950 {
 951    emit_2ub(p, 0xd9, 0xfa);
 952 }
 953
 954 void x87_fxtract( struct x86_function *p )
 955 {
 956    emit_2ub(p, 0xd9, 0xf4);
 957 }
 958
 959 /* st0 = (2^st0)-1
 960  *
 961  * Restrictions: -1.0 <= st0 <= 1.0
 962  */
 963 void x87_f2xm1( struct x86_function *p )
 964 {
 965    emit_2ub(p, 0xd9, 0xf0);
 966 }
 967
 968 /* st1 = st1 * log2(st0);
 969  * pop_stack;
 970  */
 971 void x87_fyl2x( struct x86_function *p )
 972 {
 973    emit_2ub(p, 0xd9, 0xf1);
 974 }
 975
 976 /* st1 = st1 * log2(st0 + 1.0);
 977  * pop_stack;
 978  *
 979  * A fast operation, with restrictions: -.29 < st0 < .29
 980  */
 981 void x87_fyl2xp1( struct x86_function *p )
 982 {
 983    emit_2ub(p, 0xd9, 0xf9);
 984 }
 985
 986
 987 void x87_fld( struct x86_function *p, struct x86_reg arg )
 988 {
 989    if (arg.file == file_x87)
 990       emit_2ub(p, 0xd9, 0xc0 + arg.idx);
 991    else {
 992       emit_1ub(p, 0xd9);
 993       emit_modrm_noreg(p, 0, arg);
 994    }
 995 }
 996
 997 void x87_fst( struct x86_function *p, struct x86_reg dst )
 998 {
 999    if (dst.file == file_x87)
1000       emit_2ub(p, 0xdd, 0xd0 + dst.idx);
1001    else {
1002       emit_1ub(p, 0xd9);
1003       emit_modrm_noreg(p, 2, dst);
1004    }
1005 }
1006
1007 void x87_fstp( struct x86_function *p, struct x86_reg dst )
1008 {
1009    if (dst.file == file_x87)
1010       emit_2ub(p, 0xdd, 0xd8 + dst.idx);
1011    else {
1012       emit_1ub(p, 0xd9);
1013       emit_modrm_noreg(p, 3, dst);
1014    }
1015 }
1016
1017 void x87_fcom( struct x86_function *p, struct x86_reg dst )
1018 {
1019    if (dst.file == file_x87)
1020       emit_2ub(p, 0xd8, 0xd0 + dst.idx);
1021    else {
1022       emit_1ub(p, 0xd8);
1023       emit_modrm_noreg(p, 2, dst);
1024    }
1025 }
1026
1027 void x87_fcomp( struct x86_function *p, struct x86_reg dst )
1028 {
1029    if (dst.file == file_x87)
1030       emit_2ub(p, 0xd8, 0xd8 + dst.idx);
1031    else {
1032       emit_1ub(p, 0xd8);
1033       emit_modrm_noreg(p, 3, dst);
1034    }
1035 }
1036
1037
1038 void x87_fnstsw( struct x86_function *p, struct x86_reg dst )
1039 {
1040    assert(dst.file == file_REG32);
1041
1042    if (dst.idx == reg_AX &&
1043        dst.mod == mod_REG)
1044       emit_2ub(p, 0xdf, 0xe0);
1045    else {
1046       emit_1ub(p, 0xdd);
1047       emit_modrm_noreg(p, 7, dst);
1048    }
1049 }
1050
1051
1052
1053
1054 /***********************************************************************
1055  * MMX instructions
1056  */
1057
1058 void mmx_emms( struct x86_function *p )
1059 {
1060    assert(p->need_emms);
1061    emit_2ub(p, 0x0f, 0x77);
1062    p->need_emms = 0;
1063 }
1064
1065 void mmx_packssdw( struct x86_function *p,
1066                    struct x86_reg dst,
1067                    struct x86_reg src )
1068 {
1069    assert(dst.file == file_MMX &&
1070           (src.file == file_MMX || src.mod != mod_REG));
1071
1072    p->need_emms = 1;
1073
1074    emit_2ub(p, X86_TWOB, 0x6b);
1075    emit_modrm( p, dst, src );
1076 }
1077
1078 void mmx_packuswb( struct x86_function *p,
1079                    struct x86_reg dst,
1080                    struct x86_reg src )
1081 {
1082    assert(dst.file == file_MMX &&
1083           (src.file == file_MMX || src.mod != mod_REG));
1084
1085    p->need_emms = 1;
1086
1087    emit_2ub(p, X86_TWOB, 0x67);
1088    emit_modrm( p, dst, src );
1089 }
1090
1091 void mmx_movd( struct x86_function *p,
1092                struct x86_reg dst,
1093                struct x86_reg src )
1094 {
1095    p->need_emms = 1;
1096    emit_1ub(p, X86_TWOB);
1097    emit_op_modrm( p, 0x6e, 0x7e, dst, src );
1098 }
1099
1100 void mmx_movq( struct x86_function *p,
1101                struct x86_reg dst,
1102                struct x86_reg src )
1103 {
1104    p->need_emms = 1;
1105    emit_1ub(p, X86_TWOB);
1106    emit_op_modrm( p, 0x6f, 0x7f, dst, src );
1107 }
1108
1109
1110 /***********************************************************************
1111  * Helper functions
1112  */
1113
1114
1115 /* Retreive a reference to one of the function arguments, taking into
1116  * account any push/pop activity:
1117  */
1118 struct x86_reg x86_fn_arg( struct x86_function *p,
1119                            GLuint arg )
1120 {
1121    return x86_make_disp(x86_make_reg(file_REG32, reg_SP),
1122                         p->stack_offset + arg * 4);     /* ??? */
1123 }
1124
1125
1126 void x86_init_func( struct x86_function *p )
1127 {
1128    x86_init_func_size(p, 1024);
1129 }
1130
1131 void x86_init_func_size( struct x86_function *p, GLuint code_size )
1132 {
1133    p->store = _mesa_exec_malloc(code_size);
1134    p->csr = p->store;
1135 }
1136
1137 void x86_release_func( struct x86_function *p )
1138 {
1139    _mesa_exec_free(p->store);
1140 }
1141
1142
1143 void (*x86_get_func( struct x86_function *p ))(void)
1144 {
1145    if (DISASSEM)
1146       _mesa_printf("disassemble %p %p\n", p->store, p->csr);
1147    return (void (*)(void)) (unsigned long) p->store;
1148 }
1149
1150 #else
1151
1152 void x86sse_dummy( void )
1153 {
1154 }
1155
1156 #endif