src/mesa/x86/rtasm/x86sse.c

   1 #ifdef USE_X86_ASM
   2 #if defined(__i386__) || defined(__386__)
   3
   4 #include <stdio.h>
   5
   6 #include "main/imports.h"
   7 #include "main/execmem.h"
   8 #include "x86sse.h"
   9
  10 #define DISASSEM 0
  11 #define X86_TWOB 0x0f
  12
  13 #if 0
  14 static unsigned char *cptr( void (*label)() )
  15 {
  16    return (unsigned char *)(unsigned long)label;
  17 }
  18 #endif
  19
  20
  21 static void do_realloc( struct x86_function *p )
  22 {
  23    if (p->size == 0) {
  24       p->size = 1024;
  25       p->store = _mesa_exec_malloc(p->size);
  26       p->csr = p->store;
  27    }
  28    else {
  29       unsigned used = p->csr - p->store;
  30       unsigned char *tmp = p->store;
  31       p->size *= 2;
  32       p->store = _mesa_exec_malloc(p->size);
  33       memcpy(p->store, tmp, used);
  34       p->csr = p->store + used;
  35       _mesa_exec_free(tmp);
  36    }
  37 }
  38
  39 /* Emit bytes to the instruction stream:
  40  */
  41 static unsigned char *reserve( struct x86_function *p, int bytes )
  42 {
  43    if (p->csr + bytes - p->store > p->size)
  44       do_realloc(p);
  45
  46    {
  47       unsigned char *csr = p->csr;
  48       p->csr += bytes;
  49       return csr;
  50    }
  51 }
  52
  53
  54
  55 static void emit_1b( struct x86_function *p, char b0 )
  56 {
  57    char *csr = (char *)reserve(p, 1);
  58    *csr = b0;
  59 }
  60
  61 static void emit_1i( struct x86_function *p, int i0 )
  62 {
  63    int *icsr = (int *)reserve(p, sizeof(i0));
  64    *icsr = i0;
  65 }
  66
  67 static void emit_1ub( struct x86_function *p, unsigned char b0 )
  68 {
  69    unsigned char *csr = reserve(p, 1);
  70    *csr++ = b0;
  71 }
  72
  73 static void emit_2ub( struct x86_function *p, unsigned char b0, unsigned char b1 )
  74 {
  75    unsigned char *csr = reserve(p, 2);
  76    *csr++ = b0;
  77    *csr++ = b1;
  78 }
  79
  80 static void emit_3ub( struct x86_function *p, unsigned char b0, unsigned char b1, unsigned char b2 )
  81 {
  82    unsigned char *csr = reserve(p, 3);
  83    *csr++ = b0;
  84    *csr++ = b1;
  85    *csr++ = b2;
  86 }
  87
  88
  89 /* Build a modRM byte + possible displacement.  No treatment of SIB
  90  * indexing.  BZZT - no way to encode an absolute address.
  91  */
  92 static void emit_modrm( struct x86_function *p,
  93                         struct x86_reg reg,
  94                         struct x86_reg regmem )
  95 {
  96    unsigned char val = 0;
  97
  98    assert(reg.mod == mod_REG);
  99
 100    val |= regmem.mod << 6;      /* mod field */
 101    val |= reg.idx << 3;         /* reg field */
 102    val |= regmem.idx;           /* r/m field */
 103
 104    emit_1ub(p, val);
 105
 106    /* Oh-oh we've stumbled into the SIB thing.
 107     */
 108    if (regmem.file == file_REG32 &&
 109        regmem.idx == reg_SP) {
 110       emit_1ub(p, 0x24);                /* simplistic! */
 111    }
 112
 113    switch (regmem.mod) {
 114    case mod_REG:
 115    case mod_INDIRECT:
 116       break;
 117    case mod_DISP8:
 118       emit_1b(p, regmem.disp);
 119       break;
 120    case mod_DISP32:
 121       emit_1i(p, regmem.disp);
 122       break;
 123    default:
 124       assert(0);
 125       break;
 126    }
 127 }
 128
 129
 130 static void emit_modrm_noreg( struct x86_function *p,
 131                               unsigned op,
 132                               struct x86_reg regmem )
 133 {
 134    struct x86_reg dummy = x86_make_reg(file_REG32, op);
 135    emit_modrm(p, dummy, regmem);
 136 }
 137
 138 /* Many x86 instructions have two opcodes to cope with the situations
 139  * where the destination is a register or memory reference
 140  * respectively.  This function selects the correct opcode based on
 141  * the arguments presented.
 142  */
 143 static void emit_op_modrm( struct x86_function *p,
 144                            unsigned char op_dst_is_reg,
 145                            unsigned char op_dst_is_mem,
 146                            struct x86_reg dst,
 147                            struct x86_reg src )
 148 {
 149    switch (dst.mod) {
 150    case mod_REG:
 151       emit_1ub(p, op_dst_is_reg);
 152       emit_modrm(p, dst, src);
 153       break;
 154    case mod_INDIRECT:
 155    case mod_DISP32:
 156    case mod_DISP8:
 157       assert(src.mod == mod_REG);
 158       emit_1ub(p, op_dst_is_mem);
 159       emit_modrm(p, src, dst);
 160       break;
 161    default:
 162       assert(0);
 163       break;
 164    }
 165 }
 166
 167
 168
 169
 170
 171
 172
 173 /* Create and manipulate registers and regmem values:
 174  */
 175 struct x86_reg x86_make_reg( enum x86_reg_file file,
 176                              enum x86_reg_name idx )
 177 {
 178    struct x86_reg reg;
 179
 180    reg.file = file;
 181    reg.idx = idx;
 182    reg.mod = mod_REG;
 183    reg.disp = 0;
 184
 185    return reg;
 186 }
 187
 188 struct x86_reg x86_make_disp( struct x86_reg reg,
 189                               int disp )
 190 {
 191    assert(reg.file == file_REG32);
 192
 193    if (reg.mod == mod_REG)
 194       reg.disp = disp;
 195    else
 196       reg.disp += disp;
 197
 198    if (reg.disp == 0)
 199       reg.mod = mod_INDIRECT;
 200    else if (reg.disp <= 127 && reg.disp >= -128)
 201       reg.mod = mod_DISP8;
 202    else
 203       reg.mod = mod_DISP32;
 204
 205    return reg;
 206 }
 207
 208 struct x86_reg x86_deref( struct x86_reg reg )
 209 {
 210    return x86_make_disp(reg, 0);
 211 }
 212
 213 struct x86_reg x86_get_base_reg( struct x86_reg reg )
 214 {
 215    return x86_make_reg( reg.file, reg.idx );
 216 }
 217
 218 unsigned char *x86_get_label( struct x86_function *p )
 219 {
 220    return p->csr;
 221 }
 222
 223
 224
 225 /***********************************************************************
 226  * x86 instructions
 227  */
 228
 229
 230 void x86_jcc( struct x86_function *p,
 231               enum x86_cc cc,
 232               unsigned char *label )
 233 {
 234    int offset = label - (x86_get_label(p) + 2);
 235
 236    if (offset <= 127 && offset >= -128) {
 237       emit_1ub(p, 0x70 + cc);
 238       emit_1b(p, (char) offset);
 239    }
 240    else {
 241       offset = label - (x86_get_label(p) + 6);
 242       emit_2ub(p, 0x0f, 0x80 + cc);
 243       emit_1i(p, offset);
 244    }
 245 }
 246
 247 /* Always use a 32bit offset for forward jumps:
 248  */
 249 unsigned char *x86_jcc_forward( struct x86_function *p,
 250                           enum x86_cc cc )
 251 {
 252    emit_2ub(p, 0x0f, 0x80 + cc);
 253    emit_1i(p, 0);
 254    return x86_get_label(p);
 255 }
 256
 257 unsigned char *x86_jmp_forward( struct x86_function *p)
 258 {
 259    emit_1ub(p, 0xe9);
 260    emit_1i(p, 0);
 261    return x86_get_label(p);
 262 }
 263
 264 unsigned char *x86_call_forward( struct x86_function *p)
 265 {
 266    emit_1ub(p, 0xe8);
 267    emit_1i(p, 0);
 268    return x86_get_label(p);
 269 }
 270
 271 /* Fixup offset from forward jump:
 272  */
 273 void x86_fixup_fwd_jump( struct x86_function *p,
 274                          unsigned char *fixup )
 275 {
 276    *(int *)(fixup - 4) = x86_get_label(p) - fixup;
 277 }
 278
 279 void x86_jmp( struct x86_function *p, unsigned char *label)
 280 {
 281    emit_1ub(p, 0xe9);
 282    emit_1i(p, label - x86_get_label(p) - 4);
 283 }
 284
 285 #if 0
 286 /* This doesn't work once we start reallocating & copying the
 287  * generated code on buffer fills, because the call is relative to the
 288  * current pc.
 289  */
 290 void x86_call( struct x86_function *p, void (*label)())
 291 {
 292    emit_1ub(p, 0xe8);
 293    emit_1i(p, cptr(label) - x86_get_label(p) - 4);
 294 }
 295 #else
 296 void x86_call( struct x86_function *p, struct x86_reg reg)
 297 {
 298    emit_1ub(p, 0xff);
 299    emit_modrm_noreg(p, 2, reg);
 300 }
 301 #endif
 302
 303
 304 /* michal:
 305  * Temporary. As I need immediate operands, and dont want to mess with the codegen,
 306  * I load the immediate into general purpose register and use it.
 307  */
 308 void x86_mov_reg_imm( struct x86_function *p, struct x86_reg dst, int imm )
 309 {
 310    assert(dst.mod == mod_REG);
 311    emit_1ub(p, 0xb8 + dst.idx);
 312    emit_1i(p, imm);
 313 }
 314
 315 void x86_push( struct x86_function *p,
 316                struct x86_reg reg )
 317 {
 318    assert(reg.mod == mod_REG);
 319    emit_1ub(p, 0x50 + reg.idx);
 320    p->stack_offset += 4;
 321 }
 322
 323 void x86_pop( struct x86_function *p,
 324               struct x86_reg reg )
 325 {
 326    assert(reg.mod == mod_REG);
 327    emit_1ub(p, 0x58 + reg.idx);
 328    p->stack_offset -= 4;
 329 }
 330
 331 void x86_inc( struct x86_function *p,
 332               struct x86_reg reg )
 333 {
 334    assert(reg.mod == mod_REG);
 335    emit_1ub(p, 0x40 + reg.idx);
 336 }
 337
 338 void x86_dec( struct x86_function *p,
 339               struct x86_reg reg )
 340 {
 341    assert(reg.mod == mod_REG);
 342    emit_1ub(p, 0x48 + reg.idx);
 343 }
 344
 345 void x86_ret( struct x86_function *p )
 346 {
 347    emit_1ub(p, 0xc3);
 348 }
 349
 350 void x86_sahf( struct x86_function *p )
 351 {
 352    emit_1ub(p, 0x9e);
 353 }
 354
 355 void x86_mov( struct x86_function *p,
 356               struct x86_reg dst,
 357               struct x86_reg src )
 358 {
 359    emit_op_modrm( p, 0x8b, 0x89, dst, src );
 360 }
 361
 362 void x86_xor( struct x86_function *p,
 363               struct x86_reg dst,
 364               struct x86_reg src )
 365 {
 366    emit_op_modrm( p, 0x33, 0x31, dst, src );
 367 }
 368
 369 void x86_cmp( struct x86_function *p,
 370               struct x86_reg dst,
 371               struct x86_reg src )
 372 {
 373    emit_op_modrm( p, 0x3b, 0x39, dst, src );
 374 }
 375
 376 void x86_lea( struct x86_function *p,
 377               struct x86_reg dst,
 378               struct x86_reg src )
 379 {
 380    emit_1ub(p, 0x8d);
 381    emit_modrm( p, dst, src );
 382 }
 383
 384 void x86_test( struct x86_function *p,
 385                struct x86_reg dst,
 386                struct x86_reg src )
 387 {
 388    emit_1ub(p, 0x85);
 389    emit_modrm( p, dst, src );
 390 }
 391
 392 void x86_add( struct x86_function *p,
 393                struct x86_reg dst,
 394                struct x86_reg src )
 395 {
 396    emit_op_modrm(p, 0x03, 0x01, dst, src );
 397 }
 398
 399 void x86_mul( struct x86_function *p,
 400                struct x86_reg src )
 401 {
 402    assert (src.file == file_REG32 && src.mod == mod_REG);
 403    emit_op_modrm(p, 0xf7, 0, x86_make_reg (file_REG32, reg_SP), src );
 404 }
 405
 406 void x86_sub( struct x86_function *p,
 407                struct x86_reg dst,
 408                struct x86_reg src )
 409 {
 410    emit_op_modrm(p, 0x2b, 0x29, dst, src );
 411 }
 412
 413 void x86_or( struct x86_function *p,
 414              struct x86_reg dst,
 415              struct x86_reg src )
 416 {
 417    emit_op_modrm( p, 0x0b, 0x09, dst, src );
 418 }
 419
 420 void x86_and( struct x86_function *p,
 421               struct x86_reg dst,
 422               struct x86_reg src )
 423 {
 424    emit_op_modrm( p, 0x23, 0x21, dst, src );
 425 }
 426
 427
 428
 429 /***********************************************************************
 430  * SSE instructions
 431  */
 432
 433
 434 void sse_movss( struct x86_function *p,
 435                 struct x86_reg dst,
 436                 struct x86_reg src )
 437 {
 438    emit_2ub(p, 0xF3, X86_TWOB);
 439    emit_op_modrm( p, 0x10, 0x11, dst, src );
 440 }
 441
 442 void sse_movaps( struct x86_function *p,
 443                  struct x86_reg dst,
 444                  struct x86_reg src )
 445 {
 446    emit_1ub(p, X86_TWOB);
 447    emit_op_modrm( p, 0x28, 0x29, dst, src );
 448 }
 449
 450 void sse_movups( struct x86_function *p,
 451                  struct x86_reg dst,
 452                  struct x86_reg src )
 453 {
 454    emit_1ub(p, X86_TWOB);
 455    emit_op_modrm( p, 0x10, 0x11, dst, src );
 456 }
 457
 458 void sse_movhps( struct x86_function *p,
 459                  struct x86_reg dst,
 460                  struct x86_reg src )
 461 {
 462    assert(dst.mod != mod_REG || src.mod != mod_REG);
 463    emit_1ub(p, X86_TWOB);
 464    emit_op_modrm( p, 0x16, 0x17, dst, src ); /* cf movlhps */
 465 }
 466
 467 void sse_movlps( struct x86_function *p,
 468                  struct x86_reg dst,
 469                  struct x86_reg src )
 470 {
 471    assert(dst.mod != mod_REG || src.mod != mod_REG);
 472    emit_1ub(p, X86_TWOB);
 473    emit_op_modrm( p, 0x12, 0x13, dst, src ); /* cf movhlps */
 474 }
 475
 476 void sse_maxps( struct x86_function *p,
 477                 struct x86_reg dst,
 478                 struct x86_reg src )
 479 {
 480    emit_2ub(p, X86_TWOB, 0x5F);
 481    emit_modrm( p, dst, src );
 482 }
 483
 484 void sse_maxss( struct x86_function *p,
 485                 struct x86_reg dst,
 486                 struct x86_reg src )
 487 {
 488    emit_3ub(p, 0xF3, X86_TWOB, 0x5F);
 489    emit_modrm( p, dst, src );
 490 }
 491
 492 void sse_divss( struct x86_function *p,
 493                 struct x86_reg dst,
 494                 struct x86_reg src )
 495 {
 496    emit_3ub(p, 0xF3, X86_TWOB, 0x5E);
 497    emit_modrm( p, dst, src );
 498 }
 499
 500 void sse_minps( struct x86_function *p,
 501                 struct x86_reg dst,
 502                 struct x86_reg src )
 503 {
 504    emit_2ub(p, X86_TWOB, 0x5D);
 505    emit_modrm( p, dst, src );
 506 }
 507
 508 void sse_subps( struct x86_function *p,
 509                 struct x86_reg dst,
 510                 struct x86_reg src )
 511 {
 512    emit_2ub(p, X86_TWOB, 0x5C);
 513    emit_modrm( p, dst, src );
 514 }
 515
 516 void sse_mulps( struct x86_function *p,
 517                 struct x86_reg dst,
 518                 struct x86_reg src )
 519 {
 520    emit_2ub(p, X86_TWOB, 0x59);
 521    emit_modrm( p, dst, src );
 522 }
 523
 524 void sse_mulss( struct x86_function *p,
 525                 struct x86_reg dst,
 526                 struct x86_reg src )
 527 {
 528    emit_3ub(p, 0xF3, X86_TWOB, 0x59);
 529    emit_modrm( p, dst, src );
 530 }
 531
 532 void sse_addps( struct x86_function *p,
 533                 struct x86_reg dst,
 534                 struct x86_reg src )
 535 {
 536    emit_2ub(p, X86_TWOB, 0x58);
 537    emit_modrm( p, dst, src );
 538 }
 539
 540 void sse_addss( struct x86_function *p,
 541                 struct x86_reg dst,
 542                 struct x86_reg src )
 543 {
 544    emit_3ub(p, 0xF3, X86_TWOB, 0x58);
 545    emit_modrm( p, dst, src );
 546 }
 547
 548 void sse_andnps( struct x86_function *p,
 549                  struct x86_reg dst,
 550                  struct x86_reg src )
 551 {
 552    emit_2ub(p, X86_TWOB, 0x55);
 553    emit_modrm( p, dst, src );
 554 }
 555
 556 void sse_andps( struct x86_function *p,
 557                 struct x86_reg dst,
 558                 struct x86_reg src )
 559 {
 560    emit_2ub(p, X86_TWOB, 0x54);
 561    emit_modrm( p, dst, src );
 562 }
 563
 564 void sse_rsqrtps( struct x86_function *p,
 565                   struct x86_reg dst,
 566                   struct x86_reg src )
 567 {
 568    emit_2ub(p, X86_TWOB, 0x52);
 569    emit_modrm( p, dst, src );
 570 }
 571
 572 void sse_rsqrtss( struct x86_function *p,
 573                   struct x86_reg dst,
 574                   struct x86_reg src )
 575 {
 576    emit_3ub(p, 0xF3, X86_TWOB, 0x52);
 577    emit_modrm( p, dst, src );
 578
 579 }
 580
 581 void sse_movhlps( struct x86_function *p,
 582                   struct x86_reg dst,
 583                   struct x86_reg src )
 584 {
 585    assert(dst.mod == mod_REG && src.mod == mod_REG);
 586    emit_2ub(p, X86_TWOB, 0x12);
 587    emit_modrm( p, dst, src );
 588 }
 589
 590 void sse_movlhps( struct x86_function *p,
 591                   struct x86_reg dst,
 592                   struct x86_reg src )
 593 {
 594    assert(dst.mod == mod_REG && src.mod == mod_REG);
 595    emit_2ub(p, X86_TWOB, 0x16);
 596    emit_modrm( p, dst, src );
 597 }
 598
 599 void sse_orps( struct x86_function *p,
 600                struct x86_reg dst,
 601                struct x86_reg src )
 602 {
 603    emit_2ub(p, X86_TWOB, 0x56);
 604    emit_modrm( p, dst, src );
 605 }
 606
 607 void sse_xorps( struct x86_function *p,
 608                 struct x86_reg dst,
 609                 struct x86_reg src )
 610 {
 611    emit_2ub(p, X86_TWOB, 0x57);
 612    emit_modrm( p, dst, src );
 613 }
 614
 615 void sse_cvtps2pi( struct x86_function *p,
 616                    struct x86_reg dst,
 617                    struct x86_reg src )
 618 {
 619    assert(dst.file == file_MMX &&
 620           (src.file == file_XMM || src.mod != mod_REG));
 621
 622    p->need_emms = 1;
 623
 624    emit_2ub(p, X86_TWOB, 0x2d);
 625    emit_modrm( p, dst, src );
 626 }
 627
 628
 629 /* Shufps can also be used to implement a reduced swizzle when dest ==
 630  * arg0.
 631  */
 632 void sse_shufps( struct x86_function *p,
 633                  struct x86_reg dest,
 634                  struct x86_reg arg0,
 635                  unsigned char shuf)
 636 {
 637    emit_2ub(p, X86_TWOB, 0xC6);
 638    emit_modrm(p, dest, arg0);
 639    emit_1ub(p, shuf);
 640 }
 641
 642 void sse_cmpps( struct x86_function *p,
 643                 struct x86_reg dest,
 644                 struct x86_reg arg0,
 645                 unsigned char cc)
 646 {
 647    emit_2ub(p, X86_TWOB, 0xC2);
 648    emit_modrm(p, dest, arg0);
 649    emit_1ub(p, cc);
 650 }
 651
 652 void sse_pmovmskb( struct x86_function *p,
 653                    struct x86_reg dest,
 654                    struct x86_reg src)
 655 {
 656     emit_3ub(p, 0x66, X86_TWOB, 0xD7);
 657     emit_modrm(p, dest, src);
 658 }
 659
 660 /***********************************************************************
 661  * SSE2 instructions
 662  */
 663
 664 /**
 665  * Perform a reduced swizzle:
 666  */
 667 void sse2_pshufd( struct x86_function *p,
 668                   struct x86_reg dest,
 669                   struct x86_reg arg0,
 670                   unsigned char shuf)
 671 {
 672    emit_3ub(p, 0x66, X86_TWOB, 0x70);
 673    emit_modrm(p, dest, arg0);
 674    emit_1ub(p, shuf);
 675 }
 676
 677 void sse2_cvttps2dq( struct x86_function *p,
 678                      struct x86_reg dst,
 679                      struct x86_reg src )
 680 {
 681    emit_3ub( p, 0xF3, X86_TWOB, 0x5B );
 682    emit_modrm( p, dst, src );
 683 }
 684
 685 void sse2_cvtps2dq( struct x86_function *p,
 686                     struct x86_reg dst,
 687                     struct x86_reg src )
 688 {
 689    emit_3ub(p, 0x66, X86_TWOB, 0x5B);
 690    emit_modrm( p, dst, src );
 691 }
 692
 693 void sse2_packssdw( struct x86_function *p,
 694                     struct x86_reg dst,
 695                     struct x86_reg src )
 696 {
 697    emit_3ub(p, 0x66, X86_TWOB, 0x6B);
 698    emit_modrm( p, dst, src );
 699 }
 700
 701 void sse2_packsswb( struct x86_function *p,
 702                     struct x86_reg dst,
 703                     struct x86_reg src )
 704 {
 705    emit_3ub(p, 0x66, X86_TWOB, 0x63);
 706    emit_modrm( p, dst, src );
 707 }
 708
 709 void sse2_packuswb( struct x86_function *p,
 710                     struct x86_reg dst,
 711                     struct x86_reg src )
 712 {
 713    emit_3ub(p, 0x66, X86_TWOB, 0x67);
 714    emit_modrm( p, dst, src );
 715 }
 716
 717 void sse2_rcpps( struct x86_function *p,
 718                  struct x86_reg dst,
 719                  struct x86_reg src )
 720 {
 721    emit_2ub(p, X86_TWOB, 0x53);
 722    emit_modrm( p, dst, src );
 723 }
 724
 725 void sse2_rcpss( struct x86_function *p,
 726                 struct x86_reg dst,
 727                 struct x86_reg src )
 728 {
 729    emit_3ub(p, 0xF3, X86_TWOB, 0x53);
 730    emit_modrm( p, dst, src );
 731 }
 732
 733 void sse2_movd( struct x86_function *p,
 734                 struct x86_reg dst,
 735                 struct x86_reg src )
 736 {
 737    emit_2ub(p, 0x66, X86_TWOB);
 738    emit_op_modrm( p, 0x6e, 0x7e, dst, src );
 739 }
 740
 741
 742
 743
 744 /***********************************************************************
 745  * x87 instructions
 746  */
 747 void x87_fist( struct x86_function *p, struct x86_reg dst )
 748 {
 749    emit_1ub(p, 0xdb);
 750    emit_modrm_noreg(p, 2, dst);
 751 }
 752
 753 void x87_fistp( struct x86_function *p, struct x86_reg dst )
 754 {
 755    emit_1ub(p, 0xdb);
 756    emit_modrm_noreg(p, 3, dst);
 757 }
 758
 759 void x87_fild( struct x86_function *p, struct x86_reg arg )
 760 {
 761    emit_1ub(p, 0xdf);
 762    emit_modrm_noreg(p, 0, arg);
 763 }
 764
 765 void x87_fldz( struct x86_function *p )
 766 {
 767    emit_2ub(p, 0xd9, 0xee);
 768 }
 769
 770
 771 void x87_fldcw( struct x86_function *p, struct x86_reg arg )
 772 {
 773    assert(arg.file == file_REG32);
 774    assert(arg.mod != mod_REG);
 775    emit_1ub(p, 0xd9);
 776    emit_modrm_noreg(p, 5, arg);
 777 }
 778
 779 void x87_fld1( struct x86_function *p )
 780 {
 781    emit_2ub(p, 0xd9, 0xe8);
 782 }
 783
 784 void x87_fldl2e( struct x86_function *p )
 785 {
 786    emit_2ub(p, 0xd9, 0xea);
 787 }
 788
 789 void x87_fldln2( struct x86_function *p )
 790 {
 791    emit_2ub(p, 0xd9, 0xed);
 792 }
 793
 794 void x87_fwait( struct x86_function *p )
 795 {
 796    emit_1ub(p, 0x9b);
 797 }
 798
 799 void x87_fnclex( struct x86_function *p )
 800 {
 801    emit_2ub(p, 0xdb, 0xe2);
 802 }
 803
 804 void x87_fclex( struct x86_function *p )
 805 {
 806    x87_fwait(p);
 807    x87_fnclex(p);
 808 }
 809
 810
 811 static void x87_arith_op( struct x86_function *p, struct x86_reg dst, struct x86_reg arg,
 812                           unsigned char dst0ub0,
 813                           unsigned char dst0ub1,
 814                           unsigned char arg0ub0,
 815                           unsigned char arg0ub1,
 816                           unsigned char argmem_noreg)
 817 {
 818    assert(dst.file == file_x87);
 819
 820    if (arg.file == file_x87) {
 821       if (dst.idx == 0)
 822          emit_2ub(p, dst0ub0, dst0ub1+arg.idx);
 823       else if (arg.idx == 0)
 824          emit_2ub(p, arg0ub0, arg0ub1+arg.idx);
 825       else
 826          assert(0);
 827    }
 828    else if (dst.idx == 0) {
 829       assert(arg.file == file_REG32);
 830       emit_1ub(p, 0xd8);
 831       emit_modrm_noreg(p, argmem_noreg, arg);
 832    }
 833    else
 834       assert(0);
 835 }
 836
 837 void x87_fmul( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 838 {
 839    x87_arith_op(p, dst, arg,
 840                 0xd8, 0xc8,
 841                 0xdc, 0xc8,
 842                 4);
 843 }
 844
 845 void x87_fsub( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 846 {
 847    x87_arith_op(p, dst, arg,
 848                 0xd8, 0xe0,
 849                 0xdc, 0xe8,
 850                 4);
 851 }
 852
 853 void x87_fsubr( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 854 {
 855    x87_arith_op(p, dst, arg,
 856                 0xd8, 0xe8,
 857                 0xdc, 0xe0,
 858                 5);
 859 }
 860
 861 void x87_fadd( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 862 {
 863    x87_arith_op(p, dst, arg,
 864                 0xd8, 0xc0,
 865                 0xdc, 0xc0,
 866                 0);
 867 }
 868
 869 void x87_fdiv( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 870 {
 871    x87_arith_op(p, dst, arg,
 872                 0xd8, 0xf0,
 873                 0xdc, 0xf8,
 874                 6);
 875 }
 876
 877 void x87_fdivr( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 878 {
 879    x87_arith_op(p, dst, arg,
 880                 0xd8, 0xf8,
 881                 0xdc, 0xf0,
 882                 7);
 883 }
 884
 885 void x87_fmulp( struct x86_function *p, struct x86_reg dst )
 886 {
 887    assert(dst.file == file_x87);
 888    assert(dst.idx >= 1);
 889    emit_2ub(p, 0xde, 0xc8+dst.idx);
 890 }
 891
 892 void x87_fsubp( struct x86_function *p, struct x86_reg dst )
 893 {
 894    assert(dst.file == file_x87);
 895    assert(dst.idx >= 1);
 896    emit_2ub(p, 0xde, 0xe8+dst.idx);
 897 }
 898
 899 void x87_fsubrp( struct x86_function *p, struct x86_reg dst )
 900 {
 901    assert(dst.file == file_x87);
 902    assert(dst.idx >= 1);
 903    emit_2ub(p, 0xde, 0xe0+dst.idx);
 904 }
 905
 906 void x87_faddp( struct x86_function *p, struct x86_reg dst )
 907 {
 908    assert(dst.file == file_x87);
 909    assert(dst.idx >= 1);
 910    emit_2ub(p, 0xde, 0xc0+dst.idx);
 911 }
 912
 913 void x87_fdivp( struct x86_function *p, struct x86_reg dst )
 914 {
 915    assert(dst.file == file_x87);
 916    assert(dst.idx >= 1);
 917    emit_2ub(p, 0xde, 0xf8+dst.idx);
 918 }
 919
 920 void x87_fdivrp( struct x86_function *p, struct x86_reg dst )
 921 {
 922    assert(dst.file == file_x87);
 923    assert(dst.idx >= 1);
 924    emit_2ub(p, 0xde, 0xf0+dst.idx);
 925 }
 926
 927 void x87_fucom( struct x86_function *p, struct x86_reg arg )
 928 {
 929    assert(arg.file == file_x87);
 930    emit_2ub(p, 0xdd, 0xe0+arg.idx);
 931 }
 932
 933 void x87_fucomp( struct x86_function *p, struct x86_reg arg )
 934 {
 935    assert(arg.file == file_x87);
 936    emit_2ub(p, 0xdd, 0xe8+arg.idx);
 937 }
 938
 939 void x87_fucompp( struct x86_function *p )
 940 {
 941    emit_2ub(p, 0xda, 0xe9);
 942 }
 943
 944 void x87_fxch( struct x86_function *p, struct x86_reg arg )
 945 {
 946    assert(arg.file == file_x87);
 947    emit_2ub(p, 0xd9, 0xc8+arg.idx);
 948 }
 949
 950 void x87_fabs( struct x86_function *p )
 951 {
 952    emit_2ub(p, 0xd9, 0xe1);
 953 }
 954
 955 void x87_fchs( struct x86_function *p )
 956 {
 957    emit_2ub(p, 0xd9, 0xe0);
 958 }
 959
 960 void x87_fcos( struct x86_function *p )
 961 {
 962    emit_2ub(p, 0xd9, 0xff);
 963 }
 964
 965
 966 void x87_fprndint( struct x86_function *p )
 967 {
 968    emit_2ub(p, 0xd9, 0xfc);
 969 }
 970
 971 void x87_fscale( struct x86_function *p )
 972 {
 973    emit_2ub(p, 0xd9, 0xfd);
 974 }
 975
 976 void x87_fsin( struct x86_function *p )
 977 {
 978    emit_2ub(p, 0xd9, 0xfe);
 979 }
 980
 981 void x87_fsincos( struct x86_function *p )
 982 {
 983    emit_2ub(p, 0xd9, 0xfb);
 984 }
 985
 986 void x87_fsqrt( struct x86_function *p )
 987 {
 988    emit_2ub(p, 0xd9, 0xfa);
 989 }
 990
 991 void x87_fxtract( struct x86_function *p )
 992 {
 993    emit_2ub(p, 0xd9, 0xf4);
 994 }
 995
 996 /* st0 = (2^st0)-1
 997  *
 998  * Restrictions: -1.0 <= st0 <= 1.0
 999  */
1000 void x87_f2xm1( struct x86_function *p )
1001 {
1002    emit_2ub(p, 0xd9, 0xf0);
1003 }
1004
1005 /* st1 = st1 * log2(st0);
1006  * pop_stack;
1007  */
1008 void x87_fyl2x( struct x86_function *p )
1009 {
1010    emit_2ub(p, 0xd9, 0xf1);
1011 }
1012
1013 /* st1 = st1 * log2(st0 + 1.0);
1014  * pop_stack;
1015  *
1016  * A fast operation, with restrictions: -.29 < st0 < .29
1017  */
1018 void x87_fyl2xp1( struct x86_function *p )
1019 {
1020    emit_2ub(p, 0xd9, 0xf9);
1021 }
1022
1023
1024 void x87_fld( struct x86_function *p, struct x86_reg arg )
1025 {
1026    if (arg.file == file_x87)
1027       emit_2ub(p, 0xd9, 0xc0 + arg.idx);
1028    else {
1029       emit_1ub(p, 0xd9);
1030       emit_modrm_noreg(p, 0, arg);
1031    }
1032 }
1033
1034 void x87_fst( struct x86_function *p, struct x86_reg dst )
1035 {
1036    if (dst.file == file_x87)
1037       emit_2ub(p, 0xdd, 0xd0 + dst.idx);
1038    else {
1039       emit_1ub(p, 0xd9);
1040       emit_modrm_noreg(p, 2, dst);
1041    }
1042 }
1043
1044 void x87_fstp( struct x86_function *p, struct x86_reg dst )
1045 {
1046    if (dst.file == file_x87)
1047       emit_2ub(p, 0xdd, 0xd8 + dst.idx);
1048    else {
1049       emit_1ub(p, 0xd9);
1050       emit_modrm_noreg(p, 3, dst);
1051    }
1052 }
1053
1054 void x87_fcom( struct x86_function *p, struct x86_reg dst )
1055 {
1056    if (dst.file == file_x87)
1057       emit_2ub(p, 0xd8, 0xd0 + dst.idx);
1058    else {
1059       emit_1ub(p, 0xd8);
1060       emit_modrm_noreg(p, 2, dst);
1061    }
1062 }
1063
1064 void x87_fcomp( struct x86_function *p, struct x86_reg dst )
1065 {
1066    if (dst.file == file_x87)
1067       emit_2ub(p, 0xd8, 0xd8 + dst.idx);
1068    else {
1069       emit_1ub(p, 0xd8);
1070       emit_modrm_noreg(p, 3, dst);
1071    }
1072 }
1073
1074
1075 void x87_fnstsw( struct x86_function *p, struct x86_reg dst )
1076 {
1077    assert(dst.file == file_REG32);
1078
1079    if (dst.idx == reg_AX &&
1080        dst.mod == mod_REG)
1081       emit_2ub(p, 0xdf, 0xe0);
1082    else {
1083       emit_1ub(p, 0xdd);
1084       emit_modrm_noreg(p, 7, dst);
1085    }
1086 }
1087
1088
1089
1090
1091 /***********************************************************************
1092  * MMX instructions
1093  */
1094
1095 void mmx_emms( struct x86_function *p )
1096 {
1097    assert(p->need_emms);
1098    emit_2ub(p, 0x0f, 0x77);
1099    p->need_emms = 0;
1100 }
1101
1102 void mmx_packssdw( struct x86_function *p,
1103                    struct x86_reg dst,
1104                    struct x86_reg src )
1105 {
1106    assert(dst.file == file_MMX &&
1107           (src.file == file_MMX || src.mod != mod_REG));
1108
1109    p->need_emms = 1;
1110
1111    emit_2ub(p, X86_TWOB, 0x6b);
1112    emit_modrm( p, dst, src );
1113 }
1114
1115 void mmx_packuswb( struct x86_function *p,
1116                    struct x86_reg dst,
1117                    struct x86_reg src )
1118 {
1119    assert(dst.file == file_MMX &&
1120           (src.file == file_MMX || src.mod != mod_REG));
1121
1122    p->need_emms = 1;
1123
1124    emit_2ub(p, X86_TWOB, 0x67);
1125    emit_modrm( p, dst, src );
1126 }
1127
1128 void mmx_movd( struct x86_function *p,
1129                struct x86_reg dst,
1130                struct x86_reg src )
1131 {
1132    p->need_emms = 1;
1133    emit_1ub(p, X86_TWOB);
1134    emit_op_modrm( p, 0x6e, 0x7e, dst, src );
1135 }
1136
1137 void mmx_movq( struct x86_function *p,
1138                struct x86_reg dst,
1139                struct x86_reg src )
1140 {
1141    p->need_emms = 1;
1142    emit_1ub(p, X86_TWOB);
1143    emit_op_modrm( p, 0x6f, 0x7f, dst, src );
1144 }
1145
1146
1147 /***********************************************************************
1148  * Helper functions
1149  */
1150
1151
1152 /* Retreive a reference to one of the function arguments, taking into
1153  * account any push/pop activity:
1154  */
1155 struct x86_reg x86_fn_arg( struct x86_function *p,
1156                            unsigned arg )
1157 {
1158    return x86_make_disp(x86_make_reg(file_REG32, reg_SP),
1159                         p->stack_offset + arg * 4);     /* ??? */
1160 }
1161
1162
1163 void x86_init_func( struct x86_function *p )
1164 {
1165    p->size = 0;
1166    p->store = NULL;
1167    p->csr = p->store;
1168 }
1169
1170 int x86_init_func_size( struct x86_function *p, unsigned code_size )
1171 {
1172    p->size = code_size;
1173    p->store = _mesa_exec_malloc(code_size);
1174    p->csr = p->store;
1175    return p->store != NULL;
1176 }
1177
1178 void x86_release_func( struct x86_function *p )
1179 {
1180    _mesa_exec_free(p->store);
1181    p->store = NULL;
1182    p->csr = NULL;
1183    p->size = 0;
1184 }
1185
1186
1187 void (*x86_get_func( struct x86_function *p ))(void)
1188 {
1189    if (DISASSEM && p->store)
1190       printf("disassemble %p %p\n", p->store, p->csr);
1191    return (void (*)(void)) (unsigned long) p->store;
1192 }
1193
1194 #else
1195
1196 void x86sse_dummy( void )
1197 {
1198 }
1199
1200 #endif
1201
1202 #else  /* USE_X86_ASM */
1203
1204 int x86sse_c_dummy_var; /* silence warning */
1205
1206 #endif /* USE_X86_ASM */