x86_tiles.asm

   1 bits 32
   2
   3 section .text
   4
   5 %define which   [ebp+36]        ; int which
   6 %define line    [ebp+40]        ; int line
   7 %define where   [ebp+44]        ; unsigned char *where
   8
   9 %define vram    [ebp+24]        ; unsigned char *vram
  10 %define reg     [ebp+28]        ; unsigned char reg[0x20]
  11 %define highpal [ebp+32]        ; unsigned int *highpal
  12
  13 ;%define cache_align times ($$-$) & 3 nop       ; Align to 4-byte boundary
  14 ;%define cache_align times ($$-$) & 7 nop       ; Align to 8-byte boundary
  15 %define cache_align times ($$-$) & 31 nop       ; Align to 32-byte boundary
  16
  17 global asm_tiles_init
  18 global drawtile1_solid
  19 global drawtile1
  20 global drawtile2_solid
  21 global drawtile2
  22 global drawtile3_solid
  23 global drawtile3
  24 global drawtile4_solid
  25 global drawtile4
  26
  27 ; Neat utility macro
  28 %macro triple_xor 2
  29         xor %1, %2      ; Triple XOR for a neat register exchange  ;)
  30         xor %2, %1
  31         xor %1, %2
  32 %endmacro
  33
  34 %macro blit_pixel1 1-*          ; 8bpp blitting, solid
  35         mov eax, ebx
  36         and eax, %1
  37     %if %0 > 1
  38         shr eax, byte %2
  39     %endif
  40         or eax, [esi]
  41         mov byte [edi], al
  42         inc edi
  43 %endmacro
  44
  45 %macro blit_pixel1_trans 1-*    ; 8bpp blitting, transparent
  46         mov eax, ebx
  47         and eax, %1
  48         jz %%trans
  49     %if %0 > 1
  50         shr eax, byte %2
  51     %endif
  52         or eax, [esi]
  53         mov byte [edi], al
  54     %%trans:
  55         inc edi
  56 %endmacro
  57
  58 %macro blit_pixel2 1-*          ; 16bpp blitting, solid
  59         mov eax, ebx
  60         and eax, %1
  61     %if %0 > 1
  62         shr eax, byte %2
  63     %endif
  64         lea edx, [esi+eax*4]
  65         mov eax, [edx]
  66         mov word [edi], ax
  67         add edi, byte 2
  68 %endmacro
  69
  70 %macro blit_pixel2_trans 1-*    ; 16bpp blitting, transparent
  71         mov eax, ebx
  72         and eax, %1
  73         jz %%trans
  74     %if %0 > 1
  75         shr eax, byte %2
  76     %endif
  77         lea edx, [esi+eax*4]
  78         mov eax, [edx]
  79         mov word [edi], ax
  80     %%trans:
  81         add edi, byte 2
  82 %endmacro
  83
  84 %macro blit_pixel3 1-*          ; 24bpp blitting, solid
  85         mov eax, ebx
  86         and eax, %1
  87     %if %0 > 1
  88         shr eax, byte %2
  89     %endif
  90         lea edx, [esi+eax*4+1]
  91         mov ax, word [edx]
  92         mov word [edi], ax
  93         add edi, 2
  94         dec edx
  95         mov al, byte [edx]
  96         mov byte [edi], al
  97         inc edi
  98 %endmacro
  99
 100 %macro blit_pixel3_trans 1-*    ; 24bpp blitting, transparent
 101         mov eax, ebx
 102         and eax, %1
 103         jz %%trans
 104     %if %0 > 1
 105         shr eax, byte %2
 106     %endif
 107         lea edx, [esi+eax*4+1]
 108         mov ax, word [edx]
 109         mov word [edi], ax
 110         add edi, 2
 111         dec edx
 112         mov al, byte [edx]
 113         mov byte [edi], al
 114         inc edi
 115         jmp %%next
 116     %%trans:
 117         add edi, byte 3
 118     %%next:
 119 %endmacro
 120
 121 %macro blit_pixel4 1-*          ; 32bpp blitting, solid
 122         mov eax, ebx
 123         and eax, %1
 124     %if %0 > 1
 125         shr eax, byte %2
 126     %endif
 127         lea edx, [esi+eax*4]
 128         mov eax, [edx]
 129         mov [edi], eax
 130         add edi, byte 4
 131 %endmacro
 132
 133 %macro blit_pixel4_trans 1-*    ; 32bpp blitting, transparent
 134         mov eax, ebx
 135         and eax, %1
 136         jz %%trans
 137     %if %0 > 1
 138         shr eax, byte %2
 139     %endif
 140         lea edx, [esi+eax*4]
 141         mov eax, [edx]
 142         mov [edi], eax
 143     %%trans:
 144         add edi, byte 4
 145 %endmacro
 146
 147 ; ----------------------------------------
 148 ; int _asm_tiles_init
 149 ;   (unsigned char *vram, unsigned char *reg, unsigned char *highpal)
 150 ; ----------------------------------------
 151
 152         cache_align
 153
 154 asm_tiles_init:
 155
 156         push eax
 157         push ebx
 158         push edx
 159         push esp
 160         push ebp
 161         mov ebp, esp
 162
 163         mov eax, vram
 164         mov ebx, reg
 165         mov edx, highpal
 166         mov [__vram], eax
 167         mov [__reg], ebx
 168         mov [__highpal], edx
 169
 170         pop ebp
 171         pop esp
 172         pop edx
 173         pop ebx
 174         pop eax
 175
 176         ret
 177
 178         cache_align
 179
 180 ; ----------------------------------------
 181 ; int _drawtile1_solid
 182 ;   (int which, int line, unsigned char *where)
 183 ; ----------------------------------------
 184
 185         cache_align
 186
 187 drawtile1_solid:
 188
 189         pushad
 190         mov ebp, esp
 191
 192 .setup:
 193
 194 .get_pal:
 195
 196         mov ebx, which
 197         mov esi, [__highpal]
 198         mov eax, ebx
 199         shr eax, byte 7
 200         and eax, 0xc0
 201         add esi, eax
 202         mov edi, [__reg]
 203         push esi
 204
 205 .check_y_flip:
 206
 207         mov eax, ebx
 208         xor ecx, ecx
 209         mov edx, line
 210         test eax, 0x1000
 211
 212         jz .check_interlace
 213
 214 .y_flipped:
 215
 216         xor edx, byte 7
 217
 218         cache_align
 219
 220 .check_interlace:
 221
 222         mov esi, [__reg]
 223         mov cl, [esi+12]
 224         mov esi, [__vram]
 225         and eax, 0x7ff
 226         test cl, byte 0x2
 227
 228         jz .no_interlace
 229
 230 .interlace:
 231
 232         lea edx, [edx*8]
 233         shl eax, 6
 234         jmp .check_x_flip
 235
 236         cache_align
 237
 238 .no_interlace:
 239
 240         lea edx, [edx*4]
 241         shl eax, 5
 242
 243         cache_align
 244
 245 .check_x_flip:
 246
 247         add eax, edx
 248         mov edi, where
 249         lea esi, [esi+eax]
 250         mov ebx, [esi]
 251         pop esi
 252         mov eax, which
 253         test eax, 0x800
 254
 255         jz near .x_not_flipped
 256
 257 .x_flipped:
 258
 259         blit_pixel1 0x0f000000, 24      ; pixel 8
 260         blit_pixel1 0xf0000000, 28      ; ..... 7
 261         blit_pixel1 0x000f0000, 16      ; ..... 6
 262         blit_pixel1 0x00f00000, 20      ; ..... 5
 263         blit_pixel1 0x00000f00, 8       ; ..... 4
 264         blit_pixel1 0x0000f000, 12      ; ..... 3
 265         blit_pixel1 0x0000000f          ; ..... 2
 266         blit_pixel1 0x000000f0, 4       ; ..... 1
 267
 268         jmp .cleanup
 269
 270         cache_align
 271
 272 .x_not_flipped:
 273
 274         blit_pixel1 0x000000f0, 4       ; pixel 1
 275         blit_pixel1 0x0000000f          ; ..... 2
 276         blit_pixel1 0x0000f000, 12      ; ..... 3
 277         blit_pixel1 0x00000f00, 8       ; ..... 4
 278         blit_pixel1 0x00f00000, 20      ; ..... 5
 279         blit_pixel1 0x000f0000, 16      ; ..... 6
 280         blit_pixel1 0xf0000000, 28      ; ..... 7
 281         blit_pixel1 0x0f000000, 24      ; ..... 8
 282
 283         cache_align
 284
 285 .cleanup:
 286
 287         popad
 288
 289         ret
 290
 291         cache_align
 292
 293 ; ----------------------------------------
 294
 295 drawtile1:
 296
 297         pushad
 298         mov ebp, esp
 299
 300 .setup:
 301
 302 .get_pal:
 303
 304         mov ebx, which
 305         mov esi, [__highpal]
 306         mov eax, ebx
 307         shr eax, byte 7
 308         and eax, 0xc0
 309         add esi, eax
 310         push esi
 311
 312 .check_y_flip:
 313
 314         mov eax, ebx
 315         xor ecx, ecx
 316         mov edx, line
 317         test eax, 0x1000
 318
 319         jz .check_interlace
 320
 321 .y_flipped:
 322
 323         xor edx, byte 7
 324
 325         cache_align
 326
 327 .check_interlace:
 328
 329         mov esi, [__reg]
 330         mov cl, [esi+12]
 331         mov esi, [__vram]
 332         and eax, 0x7ff
 333         test cl, byte 0x2
 334
 335         jz .no_interlace
 336
 337 .interlace:
 338
 339         lea edx, [edx*8]
 340         shl eax, 6
 341         jmp .check_x_flip
 342
 343         cache_align
 344
 345 .no_interlace:
 346
 347         lea edx, [edx*4]
 348         shl eax, 5
 349
 350         cache_align
 351
 352 .check_x_flip:
 353
 354         add eax, edx
 355         mov edi, where
 356         lea esi, [esi+eax]
 357         mov ebx, [esi]
 358         pop esi
 359         test ebx, ebx
 360
 361         jz near .cleanup                ; Don't waste time if the tile is blank!
 362
 363         mov eax, which
 364         test eax, 0x800
 365
 366         jz near .x_not_flipped
 367
 368 .x_flipped:
 369
 370         blit_pixel1_trans 0x0f000000, 24        ; pixel 8
 371         blit_pixel1_trans 0xf0000000, 28        ; ..... 7
 372         blit_pixel1_trans 0x000f0000, 16        ; ..... 6
 373         blit_pixel1_trans 0x00f00000, 20        ; ..... 5
 374         blit_pixel1_trans 0x00000f00, 8         ; ..... 4
 375         blit_pixel1_trans 0x0000f000, 12        ; ..... 3
 376         blit_pixel1_trans 0x0000000f            ; ..... 2
 377         blit_pixel1_trans 0x000000f0, 4         ; ..... 1
 378
 379         jmp .cleanup
 380
 381         cache_align
 382
 383 .x_not_flipped:
 384
 385         blit_pixel1_trans 0x000000f0, 4         ; pixel 1
 386         blit_pixel1_trans 0x0000000f            ; ..... 2
 387         blit_pixel1_trans 0x0000f000, 12        ; ..... 3
 388         blit_pixel1_trans 0x00000f00, 8         ; ..... 4
 389         blit_pixel1_trans 0x00f00000, 20        ; ..... 5
 390         blit_pixel1_trans 0x000f0000, 16        ; ..... 6
 391         blit_pixel1_trans 0xf0000000, 28        ; ..... 7
 392         blit_pixel1_trans 0x0f000000, 24        ; ..... 8
 393
 394         cache_align
 395
 396 .cleanup:
 397
 398         popad
 399
 400         ret
 401
 402         cache_align
 403
 404 ; ----------------------------------------
 405
 406         cache_align
 407
 408 drawtile2_solid:
 409
 410         pushad
 411         mov ebp, esp
 412
 413 .setup:
 414
 415 .get_pal:
 416
 417         mov ebx, which
 418         mov esi, [__highpal]
 419         mov ecx, esi
 420         mov eax, ebx
 421         shr eax, byte 7
 422         and eax, 0xc0
 423         add esi, eax
 424 ; -
 425         mov edi, [__reg]
 426         mov edx, [edi + 7]
 427         push dword [esi]
 428         and edx, 0x3f
 429         mov eax, [ecx + edx*4]
 430         mov [esi], eax
 431 ; -
 432         push esi
 433
 434 .check_y_flip:
 435
 436         mov eax, ebx
 437         xor ecx, ecx
 438         mov edx, line
 439         test eax, 0x1000
 440
 441         jz .check_interlace
 442
 443 .y_flipped:
 444
 445         xor edx, byte 7
 446
 447         cache_align
 448
 449 .check_interlace:
 450
 451         mov esi, [__reg]
 452         mov cl, [esi+12]
 453         mov esi, [__vram]
 454         and eax, 0x7ff
 455         test cl, byte 0x2
 456
 457         jz .no_interlace
 458
 459 .interlace:
 460
 461         lea edx, [edx*8]
 462         shl eax, 6
 463         jmp .check_x_flip
 464
 465         cache_align
 466
 467 .no_interlace:
 468
 469         lea edx, [edx*4]
 470         shl eax, 5
 471
 472         cache_align
 473
 474 .check_x_flip:
 475
 476         add eax, edx
 477         mov edi, where
 478         lea esi, [esi+eax]
 479         mov ebx, [esi]
 480         pop esi
 481         mov eax, which
 482         test eax, 0x800
 483
 484         jz near .x_not_flipped
 485
 486 .x_flipped:
 487
 488         blit_pixel2 0x0f000000, 24      ; pixel 8
 489         blit_pixel2 0xf0000000, 28      ; ..... 7
 490         blit_pixel2 0x000f0000, 16      ; ..... 6
 491         blit_pixel2 0x00f00000, 20      ; ..... 5
 492         blit_pixel2 0x00000f00, 8       ; ..... 4
 493         blit_pixel2 0x0000f000, 12      ; ..... 3
 494         blit_pixel2 0x0000000f          ; ..... 2
 495         blit_pixel2 0x000000f0, 4       ; ..... 1
 496
 497         jmp .cleanup
 498
 499         cache_align
 500
 501 .x_not_flipped:
 502
 503         blit_pixel2 0x000000f0, 4       ; pixel 1
 504         blit_pixel2 0x0000000f          ; ..... 2
 505         blit_pixel2 0x0000f000, 12      ; ..... 3
 506         blit_pixel2 0x00000f00, 8       ; ..... 4
 507         blit_pixel2 0x00f00000, 20      ; ..... 5
 508         blit_pixel2 0x000f0000, 16      ; ..... 6
 509         blit_pixel2 0xf0000000, 28      ; ..... 7
 510         blit_pixel2 0x0f000000, 24      ; ..... 8
 511
 512         cache_align
 513
 514 .cleanup:
 515         pop dword [esi]
 516         popad
 517         ret
 518
 519         cache_align
 520
 521 ; ----------------------------------------
 522
 523         cache_align
 524
 525 drawtile2:
 526
 527         pushad
 528         mov ebp, esp
 529
 530 .get_pal:
 531
 532         mov ebx, which
 533         mov esi, [__highpal]
 534         mov eax, ebx
 535         shr eax, 7
 536         and eax, 0xc0
 537         add esi, eax
 538         push esi
 539
 540 .check_y_flip:
 541
 542         mov eax, ebx
 543         xor ecx, ecx
 544         mov edx, line
 545         test eax, 0x1000
 546
 547         jz .check_interlace
 548
 549 .y_flipped:
 550
 551         xor edx, byte 7
 552
 553         cache_align
 554
 555 .check_interlace:
 556
 557         mov esi, [__reg]
 558         mov cl, [esi+12]
 559         mov esi, [__vram]
 560         and eax, 0x7ff
 561         test cl, byte 0x2
 562
 563         jz .no_interlace
 564
 565 .interlace:
 566
 567         lea edx, [edx*8]
 568         shl eax, 6
 569
 570         jmp .check_x_flip
 571
 572         cache_align
 573
 574 .no_interlace:
 575         lea edx, [edx*4]
 576         shl eax, 5
 577
 578         cache_align
 579
 580 .check_x_flip:
 581
 582         add eax, edx
 583         mov edi, where
 584         lea esi, [esi+eax]
 585         mov ebx, [esi]
 586         pop esi
 587         test ebx, ebx
 588
 589         jz near .cleanup                ; Don't waste time if the tile is blank!
 590
 591         mov eax, which
 592         test eax, 0x800
 593
 594         jz near .x_not_flipped
 595
 596 .x_flipped:
 597
 598         blit_pixel2_trans 0x0f000000, 24        ; pixel 8
 599         blit_pixel2_trans 0xf0000000, 28        ; ..... 7
 600         blit_pixel2_trans 0x000f0000, 16        ; ..... 6
 601         blit_pixel2_trans 0x00f00000, 20        ; ..... 5
 602         blit_pixel2_trans 0x00000f00, 8         ; ..... 4
 603         blit_pixel2_trans 0x0000f000, 12        ; ..... 3
 604         blit_pixel2_trans 0x0000000f            ; ..... 2
 605         blit_pixel2_trans 0x000000f0, 4         ; ..... 1
 606
 607         jmp .cleanup
 608
 609         cache_align
 610
 611 .x_not_flipped:
 612
 613         blit_pixel2_trans 0x000000f0, 4         ; pixel 1
 614         blit_pixel2_trans 0x0000000f            ; ..... 2
 615         blit_pixel2_trans 0x0000f000, 12        ; ..... 3
 616         blit_pixel2_trans 0x00000f00, 8         ; ..... 4
 617         blit_pixel2_trans 0x00f00000, 20        ; ..... 5
 618         blit_pixel2_trans 0x000f0000, 16        ; ..... 6
 619         blit_pixel2_trans 0xf0000000, 28        ; ..... 7
 620         blit_pixel2_trans 0x0f000000, 24        ; ..... 8
 621
 622         cache_align
 623
 624 .cleanup:
 625         popad
 626         ret
 627
 628         cache_align
 629
 630 ; ----------------------------------------
 631
 632 drawtile3_solid:
 633
 634         pushad
 635         mov ebp, esp
 636
 637 .setup:
 638
 639 .get_pal:
 640
 641         mov ebx, which
 642         mov esi, [__highpal]
 643         mov ecx, esi
 644         mov eax, ebx
 645         shr eax, byte 7
 646         and eax, 0xc0
 647         add esi, eax
 648 ; -
 649         mov edi, [__reg]
 650         mov edx, [edi + 7]
 651         push dword [esi]
 652         and edx, 0x3f
 653         mov eax, [ecx + edx*4]
 654         mov [esi], eax
 655 ; -
 656         push esi
 657
 658 .check_y_flip:
 659
 660         mov eax, ebx
 661         xor ecx, ecx
 662         mov edx, line
 663         test eax, 0x1000
 664
 665         jz .check_interlace
 666
 667 .y_flipped:
 668
 669         xor edx, byte 7
 670
 671         cache_align
 672
 673 .check_interlace:
 674
 675         mov esi, [__reg]
 676         mov cl, [esi+12]
 677         mov esi, [__vram]
 678         and eax, 0x7ff
 679         test cl, byte 0x2
 680
 681         jz .no_interlace
 682
 683 .interlace:
 684
 685         lea edx, [edx*8]
 686         shl eax, 6
 687         jmp .check_x_flip
 688
 689         cache_align
 690
 691 .no_interlace:
 692
 693         lea edx, [edx*4]
 694         shl eax, 5
 695
 696         cache_align
 697
 698 .check_x_flip:
 699
 700         add eax, edx
 701         mov edi, where
 702         lea esi, [esi+eax]
 703         mov ebx, [esi]
 704         pop esi
 705         mov eax, which
 706         test eax, 0x800
 707
 708         jz near .x_not_flipped
 709
 710 .x_flipped:
 711
 712         blit_pixel3 0x0f000000, 24      ; pixel 8
 713         blit_pixel3 0xf0000000, 28      ; ..... 7
 714         blit_pixel3 0x000f0000, 16      ; ..... 6
 715         blit_pixel3 0x00f00000, 20      ; ..... 5
 716         blit_pixel3 0x00000f00, 8       ; ..... 4
 717         blit_pixel3 0x0000f000, 12      ; ..... 3
 718         blit_pixel3 0x0000000f          ; ..... 2
 719         blit_pixel3 0x000000f0, 4       ; ..... 1
 720
 721         jmp .cleanup
 722
 723         cache_align
 724
 725 .x_not_flipped:
 726
 727         blit_pixel3 0x000000f0, 4       ; pixel 1
 728         blit_pixel3 0x0000000f          ; ..... 2
 729         blit_pixel3 0x0000f000, 12      ; ..... 3
 730         blit_pixel3 0x00000f00, 8       ; ..... 4
 731         blit_pixel3 0x00f00000, 20      ; ..... 5
 732         blit_pixel3 0x000f0000, 16      ; ..... 6
 733         blit_pixel3 0xf0000000, 28      ; ..... 7
 734         blit_pixel3 0x0f000000, 24      ; ..... 8
 735
 736         cache_align
 737
 738 .cleanup:
 739
 740         pop dword [esi]
 741         popad
 742         ret
 743
 744         cache_align
 745
 746 ; ----------------------------------------
 747
 748 drawtile3:
 749
 750         pushad
 751         mov ebp, esp
 752
 753 .setup:
 754
 755 .get_pal:
 756
 757         mov ebx, which
 758         mov esi, [__highpal]
 759         mov eax, ebx
 760         shr eax, byte 7
 761         and eax, 0xc0
 762         add esi, eax
 763         push esi
 764
 765 .check_y_flip:
 766
 767         mov eax, ebx
 768         xor ecx, ecx
 769         mov edx, line
 770         test eax, 0x1000
 771
 772         jz .check_interlace
 773
 774 .y_flipped:
 775
 776         xor edx, byte 7
 777
 778         cache_align
 779
 780 .check_interlace:
 781
 782         mov esi, [__reg]
 783         mov cl, [esi+12]
 784         mov esi, [__vram]
 785         and eax, 0x7ff
 786         test cl, byte 0x2
 787
 788         jz .no_interlace
 789
 790 .interlace:
 791
 792         lea edx, [edx*8]
 793         shl eax, 6
 794         jmp .check_x_flip
 795
 796         cache_align
 797
 798 .no_interlace:
 799
 800         lea edx, [edx*4]
 801         shl eax, 5
 802
 803         cache_align
 804
 805 .check_x_flip:
 806
 807         add eax, edx
 808         mov edi, where
 809         lea esi, [esi+eax]
 810         mov ebx, [esi]
 811         pop esi
 812         test ebx, ebx
 813
 814         jz near .cleanup                ; Don't waste time if the tile is blank!
 815
 816         mov eax, which
 817         test eax, 0x800
 818
 819         jz near .x_not_flipped
 820
 821 .x_flipped:
 822
 823         blit_pixel3_trans 0x0f000000, 24        ; pixel 8
 824         blit_pixel3_trans 0xf0000000, 28        ; ..... 7
 825         blit_pixel3_trans 0x000f0000, 16        ; ..... 6
 826         blit_pixel3_trans 0x00f00000, 20        ; ..... 5
 827         blit_pixel3_trans 0x00000f00, 8         ; ..... 4
 828         blit_pixel3_trans 0x0000f000, 12        ; ..... 3
 829         blit_pixel3_trans 0x0000000f            ; ..... 2
 830         blit_pixel3_trans 0x000000f0, 4         ; ..... 1
 831
 832         jmp .cleanup
 833
 834         cache_align
 835
 836 .x_not_flipped:
 837
 838         blit_pixel3_trans 0x000000f0, 4         ; pixel 1
 839         blit_pixel3_trans 0x0000000f            ; ..... 2
 840         blit_pixel3_trans 0x0000f000, 12        ; ..... 3
 841         blit_pixel3_trans 0x00000f00, 8         ; ..... 4
 842         blit_pixel3_trans 0x00f00000, 20        ; ..... 5
 843         blit_pixel3_trans 0x000f0000, 16        ; ..... 6
 844         blit_pixel3_trans 0xf0000000, 28        ; ..... 7
 845         blit_pixel3_trans 0x0f000000, 24        ; ..... 8
 846
 847         cache_align
 848
 849 .cleanup:
 850
 851         popad
 852
 853         ret
 854
 855         cache_align
 856
 857 ; ----------------------------------------
 858
 859 drawtile4_solid:
 860
 861         pushad
 862         mov ebp, esp
 863
 864 .setup:
 865
 866 .get_pal:
 867
 868         mov ebx, which
 869         mov esi, [__highpal]
 870         mov ecx, esi
 871         mov eax, ebx
 872         shr eax, byte 7
 873         and eax, 0xc0
 874         add esi, eax
 875 ; -
 876         mov edi, [__reg]
 877         mov edx, [edi + 7]
 878         push dword [esi]
 879         and edx, 0x3f
 880         mov eax, [ecx + edx*4]
 881         mov [esi], eax
 882 ; -
 883         push esi
 884
 885 .check_y_flip:
 886
 887         mov eax, ebx
 888         xor ecx, ecx
 889         mov edx, line
 890         test eax, 0x1000
 891
 892         jz .check_interlace
 893
 894 .y_flipped:
 895
 896         xor edx, byte 7
 897
 898         cache_align
 899
 900 .check_interlace:
 901
 902         mov esi, [__reg]
 903         mov cl, [esi+12]
 904         mov esi, [__vram]
 905         and eax, 0x7ff
 906         test cl, byte 0x2
 907
 908         jz .no_interlace
 909
 910 .interlace:
 911
 912         lea edx, [edx*8]
 913         shl eax, 6
 914         jmp .check_x_flip
 915
 916         cache_align
 917
 918 .no_interlace:
 919
 920         lea edx, [edx*4]
 921         shl eax, 5
 922
 923         cache_align
 924
 925 .check_x_flip:
 926
 927         add eax, edx
 928         mov edi, where
 929         lea esi, [esi+eax]
 930         mov ebx, [esi]
 931         pop esi
 932         mov eax, which
 933         test eax, 0x800
 934
 935         jz near .x_not_flipped
 936
 937 .x_flipped:
 938
 939         blit_pixel4 0x0f000000, 24      ; pixel 8
 940         blit_pixel4 0xf0000000, 28      ; ..... 7
 941         blit_pixel4 0x000f0000, 16      ; ..... 6
 942         blit_pixel4 0x00f00000, 20      ; ..... 5
 943         blit_pixel4 0x00000f00, 8       ; ..... 4
 944         blit_pixel4 0x0000f000, 12      ; ..... 3
 945         blit_pixel4 0x0000000f          ; ..... 2
 946         blit_pixel4 0x000000f0, 4       ; ..... 1
 947
 948         jmp .cleanup
 949
 950         cache_align
 951
 952 .x_not_flipped:
 953
 954         blit_pixel4 0x000000f0, 4       ; pixel 1
 955         blit_pixel4 0x0000000f          ; ..... 2
 956         blit_pixel4 0x0000f000, 12      ; ..... 3
 957         blit_pixel4 0x00000f00, 8       ; ..... 4
 958         blit_pixel4 0x00f00000, 20      ; ..... 5
 959         blit_pixel4 0x000f0000, 16      ; ..... 6
 960         blit_pixel4 0xf0000000, 28      ; ..... 7
 961         blit_pixel4 0x0f000000, 24      ; ..... 8
 962
 963         cache_align
 964
 965 .cleanup:
 966
 967         pop dword [esi]
 968         popad
 969         ret
 970
 971         cache_align
 972
 973 ; ----------------------------------------
 974
 975 drawtile4:
 976
 977         pushad
 978         mov ebp, esp
 979
 980 .setup:
 981
 982 .get_pal:
 983
 984         mov ebx, which
 985         mov esi, [__highpal]
 986         mov eax, ebx
 987         shr eax, byte 7
 988         and eax, 0xc0
 989         add esi, eax
 990         push esi
 991
 992 .check_y_flip:
 993
 994         mov eax, ebx
 995         xor ecx, ecx
 996         mov edx, line
 997         test eax, 0x1000
 998
 999         jz .check_interlace
1000
1001 .y_flipped:
1002
1003         xor edx, byte 7
1004
1005         cache_align
1006
1007 .check_interlace:
1008
1009         mov esi, [__reg]
1010         mov cl, [esi+12]
1011         mov esi, [__vram]
1012         and eax, 0x7ff
1013         test cl, byte 0x2
1014
1015         jz .no_interlace
1016
1017 .interlace:
1018
1019         lea edx, [edx*8]
1020         shl eax, 6
1021         jmp .check_x_flip
1022
1023         cache_align
1024
1025 .no_interlace:
1026
1027         lea edx, [edx*4]
1028         shl eax, 5
1029
1030         cache_align
1031
1032 .check_x_flip:
1033
1034         add eax, edx
1035         mov edi, where
1036         lea esi, [esi+eax]
1037         mov ebx, [esi]
1038         pop esi
1039         test ebx, ebx
1040
1041         jz near .cleanup                ; Don't waste time if the tile is blank!
1042
1043         mov eax, which
1044         test eax, 0x800
1045
1046         jz near .x_not_flipped
1047
1048 .x_flipped:
1049
1050         blit_pixel4_trans 0x0f000000, 24        ; pixel 8
1051         blit_pixel4_trans 0xf0000000, 28        ; ..... 7
1052         blit_pixel4_trans 0x000f0000, 16        ; ..... 6
1053         blit_pixel4_trans 0x00f00000, 20        ; ..... 5
1054         blit_pixel4_trans 0x00000f00, 8         ; ..... 4
1055         blit_pixel4_trans 0x0000f000, 12        ; ..... 3
1056         blit_pixel4_trans 0x0000000f            ; ..... 2
1057         blit_pixel4_trans 0x000000f0, 4         ; ..... 1
1058
1059         jmp .cleanup
1060
1061         cache_align
1062
1063 .x_not_flipped:
1064
1065         blit_pixel4_trans 0x000000f0, 4         ; pixel 1
1066         blit_pixel4_trans 0x0000000f            ; ..... 2
1067         blit_pixel4_trans 0x0000f000, 12        ; ..... 3
1068         blit_pixel4_trans 0x00000f00, 8         ; ..... 4
1069         blit_pixel4_trans 0x00f00000, 20        ; ..... 5
1070         blit_pixel4_trans 0x000f0000, 16        ; ..... 6
1071         blit_pixel4_trans 0xf0000000, 28        ; ..... 7
1072         blit_pixel4_trans 0x0f000000, 24        ; ..... 8
1073
1074         cache_align
1075
1076 .cleanup:
1077
1078         popad
1079
1080         ret
1081
1082         cache_align
1083
1084 section .data
1085
1086         __vram          dd 0
1087         __reg           dd 0
1088         __highpal       dd 0
1089
1090 ; ----------------------------------------
1091
1092 %ifdef NASM_STACK_NOEXEC
1093 section .note.GNU-stack noalloc noexec nowrite progbits
1094 %endif