src/mesa/x86/rtasm/x86sse.c

   1 #ifdef USE_X86_ASM
   2 #if defined(__i386__) || defined(__386__)
   3
   4 #include <stdio.h>
   5 #include <string.h>
   6 #include <assert.h>
   7
   8 #include "main/execmem.h"
   9 #include "x86sse.h"
  10
  11 #define DISASSEM 0
  12 #define X86_TWOB 0x0f
  13
  14 #if 0
  15 static unsigned char *cptr( void (*label)() )
  16 {
  17    return (unsigned char *)(unsigned long)label;
  18 }
  19 #endif
  20
  21
  22 static void do_realloc( struct x86_function *p )
  23 {
  24    if (p->size == 0) {
  25       p->size = 1024;
  26       p->store = _mesa_exec_malloc(p->size);
  27       p->csr = p->store;
  28    }
  29    else {
  30       unsigned used = p->csr - p->store;
  31       unsigned char *tmp = p->store;
  32       p->size *= 2;
  33       p->store = _mesa_exec_malloc(p->size);
  34       memcpy(p->store, tmp, used);
  35       p->csr = p->store + used;
  36       _mesa_exec_free(tmp);
  37    }
  38 }
  39
  40 /* Emit bytes to the instruction stream:
  41  */
  42 static unsigned char *reserve( struct x86_function *p, int bytes )
  43 {
  44    if (p->csr + bytes - p->store > p->size)
  45       do_realloc(p);
  46
  47    {
  48       unsigned char *csr = p->csr;
  49       p->csr += bytes;
  50       return csr;
  51    }
  52 }
  53
  54
  55
  56 static void emit_1b( struct x86_function *p, char b0 )
  57 {
  58    char *csr = (char *)reserve(p, 1);
  59    *csr = b0;
  60 }
  61
  62 static void emit_1i( struct x86_function *p, int i0 )
  63 {
  64    int *icsr = (int *)reserve(p, sizeof(i0));
  65    *icsr = i0;
  66 }
  67
  68 static void emit_1ub( struct x86_function *p, unsigned char b0 )
  69 {
  70    unsigned char *csr = reserve(p, 1);
  71    *csr++ = b0;
  72 }
  73
  74 static void emit_2ub( struct x86_function *p, unsigned char b0, unsigned char b1 )
  75 {
  76    unsigned char *csr = reserve(p, 2);
  77    *csr++ = b0;
  78    *csr++ = b1;
  79 }
  80
  81 static void emit_3ub( struct x86_function *p, unsigned char b0, unsigned char b1, unsigned char b2 )
  82 {
  83    unsigned char *csr = reserve(p, 3);
  84    *csr++ = b0;
  85    *csr++ = b1;
  86    *csr++ = b2;
  87 }
  88
  89
  90 /* Build a modRM byte + possible displacement.  No treatment of SIB
  91  * indexing.  BZZT - no way to encode an absolute address.
  92  */
  93 static void emit_modrm( struct x86_function *p,
  94                         struct x86_reg reg,
  95                         struct x86_reg regmem )
  96 {
  97    unsigned char val = 0;
  98
  99    assert(reg.mod == mod_REG);
 100
 101    val |= regmem.mod << 6;      /* mod field */
 102    val |= reg.idx << 3;         /* reg field */
 103    val |= regmem.idx;           /* r/m field */
 104
 105    emit_1ub(p, val);
 106
 107    /* Oh-oh we've stumbled into the SIB thing.
 108     */
 109    if (regmem.file == file_REG32 &&
 110        regmem.idx == reg_SP) {
 111       emit_1ub(p, 0x24);                /* simplistic! */
 112    }
 113
 114    switch (regmem.mod) {
 115    case mod_REG:
 116    case mod_INDIRECT:
 117       break;
 118    case mod_DISP8:
 119       emit_1b(p, regmem.disp);
 120       break;
 121    case mod_DISP32:
 122       emit_1i(p, regmem.disp);
 123       break;
 124    default:
 125       assert(0);
 126       break;
 127    }
 128 }
 129
 130
 131 static void emit_modrm_noreg( struct x86_function *p,
 132                               unsigned op,
 133                               struct x86_reg regmem )
 134 {
 135    struct x86_reg dummy = x86_make_reg(file_REG32, op);
 136    emit_modrm(p, dummy, regmem);
 137 }
 138
 139 /* Many x86 instructions have two opcodes to cope with the situations
 140  * where the destination is a register or memory reference
 141  * respectively.  This function selects the correct opcode based on
 142  * the arguments presented.
 143  */
 144 static void emit_op_modrm( struct x86_function *p,
 145                            unsigned char op_dst_is_reg,
 146                            unsigned char op_dst_is_mem,
 147                            struct x86_reg dst,
 148                            struct x86_reg src )
 149 {
 150    switch (dst.mod) {
 151    case mod_REG:
 152       emit_1ub(p, op_dst_is_reg);
 153       emit_modrm(p, dst, src);
 154       break;
 155    case mod_INDIRECT:
 156    case mod_DISP32:
 157    case mod_DISP8:
 158       assert(src.mod == mod_REG);
 159       emit_1ub(p, op_dst_is_mem);
 160       emit_modrm(p, src, dst);
 161       break;
 162    default:
 163       assert(0);
 164       break;
 165    }
 166 }
 167
 168
 169
 170
 171
 172
 173
 174 /* Create and manipulate registers and regmem values:
 175  */
 176 struct x86_reg x86_make_reg( enum x86_reg_file file,
 177                              enum x86_reg_name idx )
 178 {
 179    struct x86_reg reg;
 180
 181    reg.file = file;
 182    reg.idx = idx;
 183    reg.mod = mod_REG;
 184    reg.disp = 0;
 185
 186    return reg;
 187 }
 188
 189 struct x86_reg x86_make_disp( struct x86_reg reg,
 190                               int disp )
 191 {
 192    assert(reg.file == file_REG32);
 193
 194    if (reg.mod == mod_REG)
 195       reg.disp = disp;
 196    else
 197       reg.disp += disp;
 198
 199    if (reg.disp == 0)
 200       reg.mod = mod_INDIRECT;
 201    else if (reg.disp <= 127 && reg.disp >= -128)
 202       reg.mod = mod_DISP8;
 203    else
 204       reg.mod = mod_DISP32;
 205
 206    return reg;
 207 }
 208
 209 struct x86_reg x86_deref( struct x86_reg reg )
 210 {
 211    return x86_make_disp(reg, 0);
 212 }
 213
 214 struct x86_reg x86_get_base_reg( struct x86_reg reg )
 215 {
 216    return x86_make_reg( reg.file, reg.idx );
 217 }
 218
 219 unsigned char *x86_get_label( struct x86_function *p )
 220 {
 221    return p->csr;
 222 }
 223
 224
 225
 226 /***********************************************************************
 227  * x86 instructions
 228  */
 229
 230
 231 void x86_jcc( struct x86_function *p,
 232               enum x86_cc cc,
 233               unsigned char *label )
 234 {
 235    int offset = label - (x86_get_label(p) + 2);
 236
 237    if (offset <= 127 && offset >= -128) {
 238       emit_1ub(p, 0x70 + cc);
 239       emit_1b(p, (char) offset);
 240    }
 241    else {
 242       offset = label - (x86_get_label(p) + 6);
 243       emit_2ub(p, 0x0f, 0x80 + cc);
 244       emit_1i(p, offset);
 245    }
 246 }
 247
 248 /* Always use a 32bit offset for forward jumps:
 249  */
 250 unsigned char *x86_jcc_forward( struct x86_function *p,
 251                           enum x86_cc cc )
 252 {
 253    emit_2ub(p, 0x0f, 0x80 + cc);
 254    emit_1i(p, 0);
 255    return x86_get_label(p);
 256 }
 257
 258 unsigned char *x86_jmp_forward( struct x86_function *p)
 259 {
 260    emit_1ub(p, 0xe9);
 261    emit_1i(p, 0);
 262    return x86_get_label(p);
 263 }
 264
 265 unsigned char *x86_call_forward( struct x86_function *p)
 266 {
 267    emit_1ub(p, 0xe8);
 268    emit_1i(p, 0);
 269    return x86_get_label(p);
 270 }
 271
 272 /* Fixup offset from forward jump:
 273  */
 274 void x86_fixup_fwd_jump( struct x86_function *p,
 275                          unsigned char *fixup )
 276 {
 277    *(int *)(fixup - 4) = x86_get_label(p) - fixup;
 278 }
 279
 280 void x86_jmp( struct x86_function *p, unsigned char *label)
 281 {
 282    emit_1ub(p, 0xe9);
 283    emit_1i(p, label - x86_get_label(p) - 4);
 284 }
 285
 286 #if 0
 287 /* This doesn't work once we start reallocating & copying the
 288  * generated code on buffer fills, because the call is relative to the
 289  * current pc.
 290  */
 291 void x86_call( struct x86_function *p, void (*label)())
 292 {
 293    emit_1ub(p, 0xe8);
 294    emit_1i(p, cptr(label) - x86_get_label(p) - 4);
 295 }
 296 #else
 297 void x86_call( struct x86_function *p, struct x86_reg reg)
 298 {
 299    emit_1ub(p, 0xff);
 300    emit_modrm_noreg(p, 2, reg);
 301 }
 302 #endif
 303
 304
 305 /* michal:
 306  * Temporary. As I need immediate operands, and dont want to mess with the codegen,
 307  * I load the immediate into general purpose register and use it.
 308  */
 309 void x86_mov_reg_imm( struct x86_function *p, struct x86_reg dst, int imm )
 310 {
 311    assert(dst.mod == mod_REG);
 312    emit_1ub(p, 0xb8 + dst.idx);
 313    emit_1i(p, imm);
 314 }
 315
 316 void x86_push( struct x86_function *p,
 317                struct x86_reg reg )
 318 {
 319    assert(reg.mod == mod_REG);
 320    emit_1ub(p, 0x50 + reg.idx);
 321    p->stack_offset += 4;
 322 }
 323
 324 void x86_pop( struct x86_function *p,
 325               struct x86_reg reg )
 326 {
 327    assert(reg.mod == mod_REG);
 328    emit_1ub(p, 0x58 + reg.idx);
 329    p->stack_offset -= 4;
 330 }
 331
 332 void x86_inc( struct x86_function *p,
 333               struct x86_reg reg )
 334 {
 335    assert(reg.mod == mod_REG);
 336    emit_1ub(p, 0x40 + reg.idx);
 337 }
 338
 339 void x86_dec( struct x86_function *p,
 340               struct x86_reg reg )
 341 {
 342    assert(reg.mod == mod_REG);
 343    emit_1ub(p, 0x48 + reg.idx);
 344 }
 345
 346 void x86_ret( struct x86_function *p )
 347 {
 348    emit_1ub(p, 0xc3);
 349 }
 350
 351 void x86_sahf( struct x86_function *p )
 352 {
 353    emit_1ub(p, 0x9e);
 354 }
 355
 356 void x86_mov( struct x86_function *p,
 357               struct x86_reg dst,
 358               struct x86_reg src )
 359 {
 360    emit_op_modrm( p, 0x8b, 0x89, dst, src );
 361 }
 362
 363 void x86_xor( struct x86_function *p,
 364               struct x86_reg dst,
 365               struct x86_reg src )
 366 {
 367    emit_op_modrm( p, 0x33, 0x31, dst, src );
 368 }
 369
 370 void x86_cmp( struct x86_function *p,
 371               struct x86_reg dst,
 372               struct x86_reg src )
 373 {
 374    emit_op_modrm( p, 0x3b, 0x39, dst, src );
 375 }
 376
 377 void x86_lea( struct x86_function *p,
 378               struct x86_reg dst,
 379               struct x86_reg src )
 380 {
 381    emit_1ub(p, 0x8d);
 382    emit_modrm( p, dst, src );
 383 }
 384
 385 void x86_test( struct x86_function *p,
 386                struct x86_reg dst,
 387                struct x86_reg src )
 388 {
 389    emit_1ub(p, 0x85);
 390    emit_modrm( p, dst, src );
 391 }
 392
 393 void x86_add( struct x86_function *p,
 394                struct x86_reg dst,
 395                struct x86_reg src )
 396 {
 397    emit_op_modrm(p, 0x03, 0x01, dst, src );
 398 }
 399
 400 void x86_mul( struct x86_function *p,
 401                struct x86_reg src )
 402 {
 403    assert (src.file == file_REG32 && src.mod == mod_REG);
 404    emit_op_modrm(p, 0xf7, 0, x86_make_reg (file_REG32, reg_SP), src );
 405 }
 406
 407 void x86_sub( struct x86_function *p,
 408                struct x86_reg dst,
 409                struct x86_reg src )
 410 {
 411    emit_op_modrm(p, 0x2b, 0x29, dst, src );
 412 }
 413
 414 void x86_or( struct x86_function *p,
 415              struct x86_reg dst,
 416              struct x86_reg src )
 417 {
 418    emit_op_modrm( p, 0x0b, 0x09, dst, src );
 419 }
 420
 421 void x86_and( struct x86_function *p,
 422               struct x86_reg dst,
 423               struct x86_reg src )
 424 {
 425    emit_op_modrm( p, 0x23, 0x21, dst, src );
 426 }
 427
 428
 429
 430 /***********************************************************************
 431  * SSE instructions
 432  */
 433
 434
 435 void sse_movss( struct x86_function *p,
 436                 struct x86_reg dst,
 437                 struct x86_reg src )
 438 {
 439    emit_2ub(p, 0xF3, X86_TWOB);
 440    emit_op_modrm( p, 0x10, 0x11, dst, src );
 441 }
 442
 443 void sse_movaps( struct x86_function *p,
 444                  struct x86_reg dst,
 445                  struct x86_reg src )
 446 {
 447    emit_1ub(p, X86_TWOB);
 448    emit_op_modrm( p, 0x28, 0x29, dst, src );
 449 }
 450
 451 void sse_movups( struct x86_function *p,
 452                  struct x86_reg dst,
 453                  struct x86_reg src )
 454 {
 455    emit_1ub(p, X86_TWOB);
 456    emit_op_modrm( p, 0x10, 0x11, dst, src );
 457 }
 458
 459 void sse_movhps( struct x86_function *p,
 460                  struct x86_reg dst,
 461                  struct x86_reg src )
 462 {
 463    assert(dst.mod != mod_REG || src.mod != mod_REG);
 464    emit_1ub(p, X86_TWOB);
 465    emit_op_modrm( p, 0x16, 0x17, dst, src ); /* cf movlhps */
 466 }
 467
 468 void sse_movlps( struct x86_function *p,
 469                  struct x86_reg dst,
 470                  struct x86_reg src )
 471 {
 472    assert(dst.mod != mod_REG || src.mod != mod_REG);
 473    emit_1ub(p, X86_TWOB);
 474    emit_op_modrm( p, 0x12, 0x13, dst, src ); /* cf movhlps */
 475 }
 476
 477 void sse_maxps( struct x86_function *p,
 478                 struct x86_reg dst,
 479                 struct x86_reg src )
 480 {
 481    emit_2ub(p, X86_TWOB, 0x5F);
 482    emit_modrm( p, dst, src );
 483 }
 484
 485 void sse_maxss( struct x86_function *p,
 486                 struct x86_reg dst,
 487                 struct x86_reg src )
 488 {
 489    emit_3ub(p, 0xF3, X86_TWOB, 0x5F);
 490    emit_modrm( p, dst, src );
 491 }
 492
 493 void sse_divss( struct x86_function *p,
 494                 struct x86_reg dst,
 495                 struct x86_reg src )
 496 {
 497    emit_3ub(p, 0xF3, X86_TWOB, 0x5E);
 498    emit_modrm( p, dst, src );
 499 }
 500
 501 void sse_minps( struct x86_function *p,
 502                 struct x86_reg dst,
 503                 struct x86_reg src )
 504 {
 505    emit_2ub(p, X86_TWOB, 0x5D);
 506    emit_modrm( p, dst, src );
 507 }
 508
 509 void sse_subps( struct x86_function *p,
 510                 struct x86_reg dst,
 511                 struct x86_reg src )
 512 {
 513    emit_2ub(p, X86_TWOB, 0x5C);
 514    emit_modrm( p, dst, src );
 515 }
 516
 517 void sse_mulps( struct x86_function *p,
 518                 struct x86_reg dst,
 519                 struct x86_reg src )
 520 {
 521    emit_2ub(p, X86_TWOB, 0x59);
 522    emit_modrm( p, dst, src );
 523 }
 524
 525 void sse_mulss( struct x86_function *p,
 526                 struct x86_reg dst,
 527                 struct x86_reg src )
 528 {
 529    emit_3ub(p, 0xF3, X86_TWOB, 0x59);
 530    emit_modrm( p, dst, src );
 531 }
 532
 533 void sse_addps( struct x86_function *p,
 534                 struct x86_reg dst,
 535                 struct x86_reg src )
 536 {
 537    emit_2ub(p, X86_TWOB, 0x58);
 538    emit_modrm( p, dst, src );
 539 }
 540
 541 void sse_addss( struct x86_function *p,
 542                 struct x86_reg dst,
 543                 struct x86_reg src )
 544 {
 545    emit_3ub(p, 0xF3, X86_TWOB, 0x58);
 546    emit_modrm( p, dst, src );
 547 }
 548
 549 void sse_andnps( struct x86_function *p,
 550                  struct x86_reg dst,
 551                  struct x86_reg src )
 552 {
 553    emit_2ub(p, X86_TWOB, 0x55);
 554    emit_modrm( p, dst, src );
 555 }
 556
 557 void sse_andps( struct x86_function *p,
 558                 struct x86_reg dst,
 559                 struct x86_reg src )
 560 {
 561    emit_2ub(p, X86_TWOB, 0x54);
 562    emit_modrm( p, dst, src );
 563 }
 564
 565 void sse_rsqrtps( struct x86_function *p,
 566                   struct x86_reg dst,
 567                   struct x86_reg src )
 568 {
 569    emit_2ub(p, X86_TWOB, 0x52);
 570    emit_modrm( p, dst, src );
 571 }
 572
 573 void sse_rsqrtss( struct x86_function *p,
 574                   struct x86_reg dst,
 575                   struct x86_reg src )
 576 {
 577    emit_3ub(p, 0xF3, X86_TWOB, 0x52);
 578    emit_modrm( p, dst, src );
 579
 580 }
 581
 582 void sse_movhlps( struct x86_function *p,
 583                   struct x86_reg dst,
 584                   struct x86_reg src )
 585 {
 586    assert(dst.mod == mod_REG && src.mod == mod_REG);
 587    emit_2ub(p, X86_TWOB, 0x12);
 588    emit_modrm( p, dst, src );
 589 }
 590
 591 void sse_movlhps( struct x86_function *p,
 592                   struct x86_reg dst,
 593                   struct x86_reg src )
 594 {
 595    assert(dst.mod == mod_REG && src.mod == mod_REG);
 596    emit_2ub(p, X86_TWOB, 0x16);
 597    emit_modrm( p, dst, src );
 598 }
 599
 600 void sse_orps( struct x86_function *p,
 601                struct x86_reg dst,
 602                struct x86_reg src )
 603 {
 604    emit_2ub(p, X86_TWOB, 0x56);
 605    emit_modrm( p, dst, src );
 606 }
 607
 608 void sse_xorps( struct x86_function *p,
 609                 struct x86_reg dst,
 610                 struct x86_reg src )
 611 {
 612    emit_2ub(p, X86_TWOB, 0x57);
 613    emit_modrm( p, dst, src );
 614 }
 615
 616 void sse_cvtps2pi( struct x86_function *p,
 617                    struct x86_reg dst,
 618                    struct x86_reg src )
 619 {
 620    assert(dst.file == file_MMX &&
 621           (src.file == file_XMM || src.mod != mod_REG));
 622
 623    p->need_emms = 1;
 624
 625    emit_2ub(p, X86_TWOB, 0x2d);
 626    emit_modrm( p, dst, src );
 627 }
 628
 629
 630 /* Shufps can also be used to implement a reduced swizzle when dest ==
 631  * arg0.
 632  */
 633 void sse_shufps( struct x86_function *p,
 634                  struct x86_reg dest,
 635                  struct x86_reg arg0,
 636                  unsigned char shuf)
 637 {
 638    emit_2ub(p, X86_TWOB, 0xC6);
 639    emit_modrm(p, dest, arg0);
 640    emit_1ub(p, shuf);
 641 }
 642
 643 void sse_cmpps( struct x86_function *p,
 644                 struct x86_reg dest,
 645                 struct x86_reg arg0,
 646                 unsigned char cc)
 647 {
 648    emit_2ub(p, X86_TWOB, 0xC2);
 649    emit_modrm(p, dest, arg0);
 650    emit_1ub(p, cc);
 651 }
 652
 653 void sse_pmovmskb( struct x86_function *p,
 654                    struct x86_reg dest,
 655                    struct x86_reg src)
 656 {
 657     emit_3ub(p, 0x66, X86_TWOB, 0xD7);
 658     emit_modrm(p, dest, src);
 659 }
 660
 661 /***********************************************************************
 662  * SSE2 instructions
 663  */
 664
 665 /**
 666  * Perform a reduced swizzle:
 667  */
 668 void sse2_pshufd( struct x86_function *p,
 669                   struct x86_reg dest,
 670                   struct x86_reg arg0,
 671                   unsigned char shuf)
 672 {
 673    emit_3ub(p, 0x66, X86_TWOB, 0x70);
 674    emit_modrm(p, dest, arg0);
 675    emit_1ub(p, shuf);
 676 }
 677
 678 void sse2_cvttps2dq( struct x86_function *p,
 679                      struct x86_reg dst,
 680                      struct x86_reg src )
 681 {
 682    emit_3ub( p, 0xF3, X86_TWOB, 0x5B );
 683    emit_modrm( p, dst, src );
 684 }
 685
 686 void sse2_cvtps2dq( struct x86_function *p,
 687                     struct x86_reg dst,
 688                     struct x86_reg src )
 689 {
 690    emit_3ub(p, 0x66, X86_TWOB, 0x5B);
 691    emit_modrm( p, dst, src );
 692 }
 693
 694 void sse2_packssdw( struct x86_function *p,
 695                     struct x86_reg dst,
 696                     struct x86_reg src )
 697 {
 698    emit_3ub(p, 0x66, X86_TWOB, 0x6B);
 699    emit_modrm( p, dst, src );
 700 }
 701
 702 void sse2_packsswb( struct x86_function *p,
 703                     struct x86_reg dst,
 704                     struct x86_reg src )
 705 {
 706    emit_3ub(p, 0x66, X86_TWOB, 0x63);
 707    emit_modrm( p, dst, src );
 708 }
 709
 710 void sse2_packuswb( struct x86_function *p,
 711                     struct x86_reg dst,
 712                     struct x86_reg src )
 713 {
 714    emit_3ub(p, 0x66, X86_TWOB, 0x67);
 715    emit_modrm( p, dst, src );
 716 }
 717
 718 void sse2_rcpps( struct x86_function *p,
 719                  struct x86_reg dst,
 720                  struct x86_reg src )
 721 {
 722    emit_2ub(p, X86_TWOB, 0x53);
 723    emit_modrm( p, dst, src );
 724 }
 725
 726 void sse2_rcpss( struct x86_function *p,
 727                 struct x86_reg dst,
 728                 struct x86_reg src )
 729 {
 730    emit_3ub(p, 0xF3, X86_TWOB, 0x53);
 731    emit_modrm( p, dst, src );
 732 }
 733
 734 void sse2_movd( struct x86_function *p,
 735                 struct x86_reg dst,
 736                 struct x86_reg src )
 737 {
 738    emit_2ub(p, 0x66, X86_TWOB);
 739    emit_op_modrm( p, 0x6e, 0x7e, dst, src );
 740 }
 741
 742
 743
 744
 745 /***********************************************************************
 746  * x87 instructions
 747  */
 748 void x87_fist( struct x86_function *p, struct x86_reg dst )
 749 {
 750    emit_1ub(p, 0xdb);
 751    emit_modrm_noreg(p, 2, dst);
 752 }
 753
 754 void x87_fistp( struct x86_function *p, struct x86_reg dst )
 755 {
 756    emit_1ub(p, 0xdb);
 757    emit_modrm_noreg(p, 3, dst);
 758 }
 759
 760 void x87_fild( struct x86_function *p, struct x86_reg arg )
 761 {
 762    emit_1ub(p, 0xdf);
 763    emit_modrm_noreg(p, 0, arg);
 764 }
 765
 766 void x87_fldz( struct x86_function *p )
 767 {
 768    emit_2ub(p, 0xd9, 0xee);
 769 }
 770
 771
 772 void x87_fldcw( struct x86_function *p, struct x86_reg arg )
 773 {
 774    assert(arg.file == file_REG32);
 775    assert(arg.mod != mod_REG);
 776    emit_1ub(p, 0xd9);
 777    emit_modrm_noreg(p, 5, arg);
 778 }
 779
 780 void x87_fld1( struct x86_function *p )
 781 {
 782    emit_2ub(p, 0xd9, 0xe8);
 783 }
 784
 785 void x87_fldl2e( struct x86_function *p )
 786 {
 787    emit_2ub(p, 0xd9, 0xea);
 788 }
 789
 790 void x87_fldln2( struct x86_function *p )
 791 {
 792    emit_2ub(p, 0xd9, 0xed);
 793 }
 794
 795 void x87_fwait( struct x86_function *p )
 796 {
 797    emit_1ub(p, 0x9b);
 798 }
 799
 800 void x87_fnclex( struct x86_function *p )
 801 {
 802    emit_2ub(p, 0xdb, 0xe2);
 803 }
 804
 805 void x87_fclex( struct x86_function *p )
 806 {
 807    x87_fwait(p);
 808    x87_fnclex(p);
 809 }
 810
 811
 812 static void x87_arith_op( struct x86_function *p, struct x86_reg dst, struct x86_reg arg,
 813                           unsigned char dst0ub0,
 814                           unsigned char dst0ub1,
 815                           unsigned char arg0ub0,
 816                           unsigned char arg0ub1,
 817                           unsigned char argmem_noreg)
 818 {
 819    assert(dst.file == file_x87);
 820
 821    if (arg.file == file_x87) {
 822       if (dst.idx == 0)
 823          emit_2ub(p, dst0ub0, dst0ub1+arg.idx);
 824       else if (arg.idx == 0)
 825          emit_2ub(p, arg0ub0, arg0ub1+arg.idx);
 826       else
 827          assert(0);
 828    }
 829    else if (dst.idx == 0) {
 830       assert(arg.file == file_REG32);
 831       emit_1ub(p, 0xd8);
 832       emit_modrm_noreg(p, argmem_noreg, arg);
 833    }
 834    else
 835       assert(0);
 836 }
 837
 838 void x87_fmul( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 839 {
 840    x87_arith_op(p, dst, arg,
 841                 0xd8, 0xc8,
 842                 0xdc, 0xc8,
 843                 4);
 844 }
 845
 846 void x87_fsub( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 847 {
 848    x87_arith_op(p, dst, arg,
 849                 0xd8, 0xe0,
 850                 0xdc, 0xe8,
 851                 4);
 852 }
 853
 854 void x87_fsubr( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 855 {
 856    x87_arith_op(p, dst, arg,
 857                 0xd8, 0xe8,
 858                 0xdc, 0xe0,
 859                 5);
 860 }
 861
 862 void x87_fadd( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 863 {
 864    x87_arith_op(p, dst, arg,
 865                 0xd8, 0xc0,
 866                 0xdc, 0xc0,
 867                 0);
 868 }
 869
 870 void x87_fdiv( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 871 {
 872    x87_arith_op(p, dst, arg,
 873                 0xd8, 0xf0,
 874                 0xdc, 0xf8,
 875                 6);
 876 }
 877
 878 void x87_fdivr( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
 879 {
 880    x87_arith_op(p, dst, arg,
 881                 0xd8, 0xf8,
 882                 0xdc, 0xf0,
 883                 7);
 884 }
 885
 886 void x87_fmulp( struct x86_function *p, struct x86_reg dst )
 887 {
 888    assert(dst.file == file_x87);
 889    assert(dst.idx >= 1);
 890    emit_2ub(p, 0xde, 0xc8+dst.idx);
 891 }
 892
 893 void x87_fsubp( struct x86_function *p, struct x86_reg dst )
 894 {
 895    assert(dst.file == file_x87);
 896    assert(dst.idx >= 1);
 897    emit_2ub(p, 0xde, 0xe8+dst.idx);
 898 }
 899
 900 void x87_fsubrp( struct x86_function *p, struct x86_reg dst )
 901 {
 902    assert(dst.file == file_x87);
 903    assert(dst.idx >= 1);
 904    emit_2ub(p, 0xde, 0xe0+dst.idx);
 905 }
 906
 907 void x87_faddp( struct x86_function *p, struct x86_reg dst )
 908 {
 909    assert(dst.file == file_x87);
 910    assert(dst.idx >= 1);
 911    emit_2ub(p, 0xde, 0xc0+dst.idx);
 912 }
 913
 914 void x87_fdivp( struct x86_function *p, struct x86_reg dst )
 915 {
 916    assert(dst.file == file_x87);
 917    assert(dst.idx >= 1);
 918    emit_2ub(p, 0xde, 0xf8+dst.idx);
 919 }
 920
 921 void x87_fdivrp( struct x86_function *p, struct x86_reg dst )
 922 {
 923    assert(dst.file == file_x87);
 924    assert(dst.idx >= 1);
 925    emit_2ub(p, 0xde, 0xf0+dst.idx);
 926 }
 927
 928 void x87_fucom( struct x86_function *p, struct x86_reg arg )
 929 {
 930    assert(arg.file == file_x87);
 931    emit_2ub(p, 0xdd, 0xe0+arg.idx);
 932 }
 933
 934 void x87_fucomp( struct x86_function *p, struct x86_reg arg )
 935 {
 936    assert(arg.file == file_x87);
 937    emit_2ub(p, 0xdd, 0xe8+arg.idx);
 938 }
 939
 940 void x87_fucompp( struct x86_function *p )
 941 {
 942    emit_2ub(p, 0xda, 0xe9);
 943 }
 944
 945 void x87_fxch( struct x86_function *p, struct x86_reg arg )
 946 {
 947    assert(arg.file == file_x87);
 948    emit_2ub(p, 0xd9, 0xc8+arg.idx);
 949 }
 950
 951 void x87_fabs( struct x86_function *p )
 952 {
 953    emit_2ub(p, 0xd9, 0xe1);
 954 }
 955
 956 void x87_fchs( struct x86_function *p )
 957 {
 958    emit_2ub(p, 0xd9, 0xe0);
 959 }
 960
 961 void x87_fcos( struct x86_function *p )
 962 {
 963    emit_2ub(p, 0xd9, 0xff);
 964 }
 965
 966
 967 void x87_fprndint( struct x86_function *p )
 968 {
 969    emit_2ub(p, 0xd9, 0xfc);
 970 }
 971
 972 void x87_fscale( struct x86_function *p )
 973 {
 974    emit_2ub(p, 0xd9, 0xfd);
 975 }
 976
 977 void x87_fsin( struct x86_function *p )
 978 {
 979    emit_2ub(p, 0xd9, 0xfe);
 980 }
 981
 982 void x87_fsincos( struct x86_function *p )
 983 {
 984    emit_2ub(p, 0xd9, 0xfb);
 985 }
 986
 987 void x87_fsqrt( struct x86_function *p )
 988 {
 989    emit_2ub(p, 0xd9, 0xfa);
 990 }
 991
 992 void x87_fxtract( struct x86_function *p )
 993 {
 994    emit_2ub(p, 0xd9, 0xf4);
 995 }
 996
 997 /* st0 = (2^st0)-1
 998  *
 999  * Restrictions: -1.0 <= st0 <= 1.0
1000  */
1001 void x87_f2xm1( struct x86_function *p )
1002 {
1003    emit_2ub(p, 0xd9, 0xf0);
1004 }
1005
1006 /* st1 = st1 * log2(st0);
1007  * pop_stack;
1008  */
1009 void x87_fyl2x( struct x86_function *p )
1010 {
1011    emit_2ub(p, 0xd9, 0xf1);
1012 }
1013
1014 /* st1 = st1 * log2(st0 + 1.0);
1015  * pop_stack;
1016  *
1017  * A fast operation, with restrictions: -.29 < st0 < .29
1018  */
1019 void x87_fyl2xp1( struct x86_function *p )
1020 {
1021    emit_2ub(p, 0xd9, 0xf9);
1022 }
1023
1024
1025 void x87_fld( struct x86_function *p, struct x86_reg arg )
1026 {
1027    if (arg.file == file_x87)
1028       emit_2ub(p, 0xd9, 0xc0 + arg.idx);
1029    else {
1030       emit_1ub(p, 0xd9);
1031       emit_modrm_noreg(p, 0, arg);
1032    }
1033 }
1034
1035 void x87_fst( struct x86_function *p, struct x86_reg dst )
1036 {
1037    if (dst.file == file_x87)
1038       emit_2ub(p, 0xdd, 0xd0 + dst.idx);
1039    else {
1040       emit_1ub(p, 0xd9);
1041       emit_modrm_noreg(p, 2, dst);
1042    }
1043 }
1044
1045 void x87_fstp( struct x86_function *p, struct x86_reg dst )
1046 {
1047    if (dst.file == file_x87)
1048       emit_2ub(p, 0xdd, 0xd8 + dst.idx);
1049    else {
1050       emit_1ub(p, 0xd9);
1051       emit_modrm_noreg(p, 3, dst);
1052    }
1053 }
1054
1055 void x87_fcom( struct x86_function *p, struct x86_reg dst )
1056 {
1057    if (dst.file == file_x87)
1058       emit_2ub(p, 0xd8, 0xd0 + dst.idx);
1059    else {
1060       emit_1ub(p, 0xd8);
1061       emit_modrm_noreg(p, 2, dst);
1062    }
1063 }
1064
1065 void x87_fcomp( struct x86_function *p, struct x86_reg dst )
1066 {
1067    if (dst.file == file_x87)
1068       emit_2ub(p, 0xd8, 0xd8 + dst.idx);
1069    else {
1070       emit_1ub(p, 0xd8);
1071       emit_modrm_noreg(p, 3, dst);
1072    }
1073 }
1074
1075
1076 void x87_fnstsw( struct x86_function *p, struct x86_reg dst )
1077 {
1078    assert(dst.file == file_REG32);
1079
1080    if (dst.idx == reg_AX &&
1081        dst.mod == mod_REG)
1082       emit_2ub(p, 0xdf, 0xe0);
1083    else {
1084       emit_1ub(p, 0xdd);
1085       emit_modrm_noreg(p, 7, dst);
1086    }
1087 }
1088
1089
1090
1091
1092 /***********************************************************************
1093  * MMX instructions
1094  */
1095
1096 void mmx_emms( struct x86_function *p )
1097 {
1098    assert(p->need_emms);
1099    emit_2ub(p, 0x0f, 0x77);
1100    p->need_emms = 0;
1101 }
1102
1103 void mmx_packssdw( struct x86_function *p,
1104                    struct x86_reg dst,
1105                    struct x86_reg src )
1106 {
1107    assert(dst.file == file_MMX &&
1108           (src.file == file_MMX || src.mod != mod_REG));
1109
1110    p->need_emms = 1;
1111
1112    emit_2ub(p, X86_TWOB, 0x6b);
1113    emit_modrm( p, dst, src );
1114 }
1115
1116 void mmx_packuswb( struct x86_function *p,
1117                    struct x86_reg dst,
1118                    struct x86_reg src )
1119 {
1120    assert(dst.file == file_MMX &&
1121           (src.file == file_MMX || src.mod != mod_REG));
1122
1123    p->need_emms = 1;
1124
1125    emit_2ub(p, X86_TWOB, 0x67);
1126    emit_modrm( p, dst, src );
1127 }
1128
1129 void mmx_movd( struct x86_function *p,
1130                struct x86_reg dst,
1131                struct x86_reg src )
1132 {
1133    p->need_emms = 1;
1134    emit_1ub(p, X86_TWOB);
1135    emit_op_modrm( p, 0x6e, 0x7e, dst, src );
1136 }
1137
1138 void mmx_movq( struct x86_function *p,
1139                struct x86_reg dst,
1140                struct x86_reg src )
1141 {
1142    p->need_emms = 1;
1143    emit_1ub(p, X86_TWOB);
1144    emit_op_modrm( p, 0x6f, 0x7f, dst, src );
1145 }
1146
1147
1148 /***********************************************************************
1149  * Helper functions
1150  */
1151
1152
1153 /* Retreive a reference to one of the function arguments, taking into
1154  * account any push/pop activity:
1155  */
1156 struct x86_reg x86_fn_arg( struct x86_function *p,
1157                            unsigned arg )
1158 {
1159    return x86_make_disp(x86_make_reg(file_REG32, reg_SP),
1160                         p->stack_offset + arg * 4);     /* ??? */
1161 }
1162
1163
1164 void x86_init_func( struct x86_function *p )
1165 {
1166    p->size = 0;
1167    p->store = NULL;
1168    p->csr = p->store;
1169 }
1170
1171 int x86_init_func_size( struct x86_function *p, unsigned code_size )
1172 {
1173    p->size = code_size;
1174    p->store = _mesa_exec_malloc(code_size);
1175    p->csr = p->store;
1176    return p->store != NULL;
1177 }
1178
1179 void x86_release_func( struct x86_function *p )
1180 {
1181    _mesa_exec_free(p->store);
1182    p->store = NULL;
1183    p->csr = NULL;
1184    p->size = 0;
1185 }
1186
1187
1188 void (*x86_get_func( struct x86_function *p ))(void)
1189 {
1190    if (DISASSEM && p->store)
1191       printf("disassemble %p %p\n", p->store, p->csr);
1192    return (void (*)(void)) (unsigned long) p->store;
1193 }
1194
1195 #else
1196
1197 void x86sse_dummy( void )
1198 {
1199 }
1200
1201 #endif
1202
1203 #else  /* USE_X86_ASM */
1204
1205 int x86sse_c_dummy_var; /* silence warning */
1206
1207 #endif /* USE_X86_ASM */