Shamusworld >> Repos - virtualjaguar/blob - src/objectp.cpp

   1 //
   2 // Object Processor
   3 //
   4 // by cal2
   5 // GCC/SDL port by Niels Wagenaar (Linux/WIN32) and Caz (BeOS)
   6 // Cleanups/fixes/rewrites by James L. Hammons
   7 //
   8
   9 #include <stdio.h>
  10 #include <stdlib.h>
  11 #include <string.h>
  12 #include "jaguar.h"
  13
  14 //#define OP_DEBUG
  15 //#define OP_DEBUG_BMP
  16
  17 #define BLEND_Y(dst, src)       op_blend_y[(((uint16)dst<<8)) | ((uint16)(src))]
  18 #define BLEND_CR(dst, src)      op_blend_cr[(((uint16)dst)<<8) | ((uint16)(src))]
  19
  20 #define OBJECT_TYPE_BITMAP      0                       // 000
  21 #define OBJECT_TYPE_SCALE       1                       // 001
  22 #define OBJECT_TYPE_GPU         2                       // 010
  23 #define OBJECT_TYPE_BRANCH      3                       // 011
  24 #define OBJECT_TYPE_STOP        4                       // 100
  25
  26 #define CONDITION_EQUAL                         0
  27 #define CONDITION_LESS_THAN                     1
  28 #define CONDITION_GREATER_THAN          2
  29 #define CONDITION_OP_FLAG_SET           3
  30 #define CONDITION_SECOND_HALF_LINE      4
  31
  32 #define OPFLAG_RELEASE          8                       // Bus release bit
  33 #define OPFLAG_TRANS            4                       // Transparency bit
  34 #define OPFLAG_RMW                      2                       // Read-Modify-Write bit
  35 #define OPFLAG_REFLECT          1                       // Horizontal mirror bit
  36
  37 // Private function prototypes
  38
  39 void OPProcessFixedBitmap(int scanline, uint64 p0, uint64 p1, bool render);
  40 void OPProcessScaledBitmap(int scanline, uint64 p0, uint64 p1, uint64 p2, bool render);
  41 void DumpScaledObject(uint64 p0, uint64 p1, uint64 p2);
  42 void DumpFixedObject(uint64 p0, uint64 p1);
  43 uint64 op_load_phrase(uint32 offset);
  44
  45 // External global variables
  46
  47 extern uint32 jaguar_mainRom_crc32;
  48
  49 // Local global variables
  50
  51 static uint8 * op_blend_y;
  52 static uint8 * op_blend_cr;
  53 // There may be a problem with this "RAM" overlapping (and thus being independent of)
  54 // some of the regular TOM RAM...
  55 static uint8 objectp_ram[0x40];                 // This is based at $F00000
  56 uint8 objectp_running;
  57 bool objectp_stop_reading_list;
  58
  59 static uint8 op_bitmap_bit_depth[8] = { 1, 2, 4, 8, 16, 24, 32, 0 };
  60 //static uint32 op_bitmap_bit_size[8] =
  61 //      { (uint32)(0.125*65536), (uint32)(0.25*65536), (uint32)(0.5*65536), (uint32)(1*65536),
  62 //        (uint32)(2*65536),     (uint32)(1*65536),    (uint32)(1*65536),   (uint32)(1*65536) };
  63 static uint32 op_pointer;
  64
  65 int32 phraseWidthToPixels[8] = { 64, 32, 16, 8, 4, 2, 0, 0 };
  66
  67
  68 //
  69 // Object Processor initialization
  70 //
  71 void op_init(void)
  72 {
  73         // Blend tables (64K each)
  74         memory_malloc_secure((void **)&op_blend_y, 0x10000, "Jaguar Object processor Y blend lookup table");
  75         memory_malloc_secure((void **)&op_blend_cr, 0x10000, "Jaguar Object processor CR blend lookup table");
  76
  77         // Here we calculate the saturating blend of a signed 4-bit value and an
  78         // existing Cyan/Red value as well as a signed 8-bit value and an existing intensity...
  79         // Note: CRY is 4 bits Cyan, 4 bits Red, 16 bits intensitY
  80         for(int i=0; i<256*256; i++)
  81         {
  82                 int y = (i >> 8) & 0xFF;
  83                 int dy = (INT8)i;                                       // Sign extend the Y index
  84                 int c1 = (i >> 8) & 0x0F;
  85                 int dc1 = (INT8)(i << 4) >> 4;          // Sign extend the R index
  86                 int c2 = (i >> 12) & 0x0F;
  87                 int dc2 = (INT8)(i & 0xF0) >> 4;        // Sign extend the C index
  88
  89                 y += dy;
  90                 if (y < 0)
  91                         y = 0;
  92                 else if (y > 0xFF)
  93                         y = 0xFF;
  94                 op_blend_y[i] = y;
  95
  96                 c1 += dc1;
  97                 if (c1 < 0)
  98                         c1 = 0;
  99                 else if (c1 > 0x0F)
 100                         c1 = 0x0F;
 101                 c2 += dc2;
 102
 103                 if (c2 < 0)
 104                         c2 = 0;
 105                 else if (c2 > 0x0F)
 106                         c2 = 0x0F;
 107                 op_blend_cr[i] = (c2 << 4) | c1;
 108         }
 109
 110         op_reset();
 111 }
 112
 113 //
 114 // Object Processor reset
 115 //
 116 void op_reset(void)
 117 {
 118         memset(objectp_ram, 0x00, 0x40);
 119         objectp_running = 0;
 120 }
 121
 122 void op_done(void)
 123 {
 124         char * opType[8] =
 125         { "(BITMAP)", "(SCALED BITMAP)", "(GPU INT)", "(BRANCH)", "(STOP)", "???", "???", "???" };
 126         char * ccType[8] =
 127                 { "\"==\"", "\"<\"", "\">\"", "(opflag set)", "(second half line)", "?", "?", "?" };
 128
 129         uint32 olp = op_get_list_pointer();
 130         WriteLog("OP: OLP = %08X\n", olp);
 131         WriteLog("OP: Phrase dump\n    ----------\n");
 132         for(uint32 i=0; i<0x100; i+=8)
 133         {
 134                 uint32 hi = JaguarReadLong(olp + i, OP), lo = JaguarReadLong(olp + i + 4, OP);
 135                 WriteLog("\t%08X: %08X %08X %s", olp + i, hi, lo, opType[lo & 0x07]);
 136                 if ((lo & 0x07) == 3)
 137                 {
 138                         uint16 ypos = (lo >> 3) & 0x7FF;
 139                         uint8  cc   = (lo >> 14) & 0x03;
 140                         uint32 link = ((hi << 11) | (lo >> 21)) & 0x3FFFF8;
 141                         WriteLog(" YPOS=%u, CC=%s, link=%08X", ypos, ccType[cc], link);
 142                 }
 143                 WriteLog("\n");
 144                 if ((lo & 0x07) == 0)
 145                         DumpFixedObject(op_load_phrase(olp+i), op_load_phrase(olp+i+8));
 146                 if ((lo & 0x07) == 1)
 147                         DumpScaledObject(op_load_phrase(olp+i), op_load_phrase(olp+i+8), op_load_phrase(olp+i+16));
 148         }
 149         WriteLog("\n");
 150 }
 151
 152 //
 153 // Object Processor memory access
 154 // Memory range: F00010 - F00027
 155 //
 156 uint8 OPReadByte(uint32 offset, uint32 who/*=UNKNOWN*/)
 157 {
 158         offset &= 0x3F;
 159         return objectp_ram[offset];
 160 }
 161
 162 uint16 OPReadWord(uint32 offset, uint32 who/*=UNKNOWN*/)
 163 {
 164         offset &= 0x3F;
 165         return GET16(objectp_ram, offset);
 166 }
 167
 168 //      F00010-F00017   R     xxxxxxxx xxxxxxxx   OB - current object code from the graphics processor
 169 //      F00020-F00023     W   xxxxxxxx xxxxxxxx   OLP - start of the object list
 170 //      F00026            W   -------- -------x   OBF - object processor flag
 171
 172 void OPWriteByte(uint32 offset, uint8 data, uint32 who/*=UNKNOWN*/)
 173 {
 174         offset &= 0x3F;
 175         objectp_ram[offset] = data;
 176 }
 177
 178 void OPWriteWord(uint32 offset, uint16 data, uint32 who/*=UNKNOWN*/)
 179 {
 180         offset &= 0x3F;
 181         SET16(objectp_ram, offset, data);
 182
 183 /*if (offset == 0x20)
 184 WriteLog("OP: Setting lo list pointer: %04X\n", data);
 185 if (offset == 0x22)
 186 WriteLog("OP: Setting hi list pointer: %04X\n", data);//*/
 187 }
 188
 189 uint32 op_get_list_pointer(void)
 190 {
 191         // Note: This register is LO / HI WORD, hence the funky look of this...
 192 //      return (objectp_ram[0x22] << 24) | (objectp_ram[0x23] << 16) | (objectp_ram[0x20] << 8) | objectp_ram[0x21];
 193         return GET16(objectp_ram, 0x20) | (GET16(objectp_ram, 0x22) << 16);
 194 }
 195
 196 // This is WRONG, since the OBF is only 16 bits wide!!! [FIXED]
 197
 198 uint32 op_get_status_register(void)
 199 {
 200 //      return (objectp_ram[0x26] << 24) | (objectp_ram[0x27] << 16) | (objectp_ram[0x28] << 8) | objectp_ram[0x29];
 201 //      return GET32(objectp_ram, 0x26);
 202         return GET16(objectp_ram, 0x26);
 203 }
 204
 205 // This is WRONG, since the OBF is only 16 bits wide!!! [FIXED]
 206
 207 void op_set_status_register(uint32 data)
 208 {
 209 /*      objectp_ram[0x26] = (data & 0xFF000000) >> 24;
 210         objectp_ram[0x27] = (data & 0x00FF0000) >> 16;
 211         objectp_ram[0x28] = (data & 0x0000FF00) >> 8;
 212         objectp_ram[0x29] |= (data & 0xFE);*/
 213         objectp_ram[0x26] = (data & 0x0000FF00) >> 8;
 214         objectp_ram[0x27] |= (data & 0xFE);
 215 }
 216
 217 void op_set_current_object(uint64 object)
 218 {
 219 //Not sure this is right... Wouldn't it just be stored 64 bit BE?
 220         // Stored as least significant 32 bits first, ms32 last in big endian
 221 /*      objectp_ram[0x13] = object & 0xFF; object >>= 8;
 222         objectp_ram[0x12] = object & 0xFF; object >>= 8;
 223         objectp_ram[0x11] = object & 0xFF; object >>= 8;
 224         objectp_ram[0x10] = object & 0xFF; object >>= 8;
 225
 226         objectp_ram[0x17] = object & 0xFF; object >>= 8;
 227         objectp_ram[0x16] = object & 0xFF; object >>= 8;
 228         objectp_ram[0x15] = object & 0xFF; object >>= 8;
 229         objectp_ram[0x14] = object & 0xFF;*/
 230 // Let's try regular good old big endian...
 231         objectp_ram[0x17] = object & 0xFF; object >>= 8;
 232         objectp_ram[0x16] = object & 0xFF; object >>= 8;
 233         objectp_ram[0x15] = object & 0xFF; object >>= 8;
 234         objectp_ram[0x14] = object & 0xFF; object >>= 8;
 235
 236         objectp_ram[0x13] = object & 0xFF; object >>= 8;
 237         objectp_ram[0x12] = object & 0xFF; object >>= 8;
 238         objectp_ram[0x11] = object & 0xFF; object >>= 8;
 239         objectp_ram[0x10] = object & 0xFF;
 240 }
 241
 242 uint64 op_load_phrase(uint32 offset)
 243 {
 244         offset &= ~0x07;                                                // 8 byte alignment
 245         return ((uint64)JaguarReadLong(offset, OP) << 32) | (uint64)JaguarReadLong(offset+4, OP);
 246 }
 247
 248 void OPStorePhrase(uint32 offset, uint64 p)
 249 {
 250         offset &= ~0x07;                                                // 8 byte alignment
 251         JaguarWriteLong(offset, p >> 32, OP);
 252         JaguarWriteLong(offset + 4, p & 0xFFFFFFFF, OP);
 253 }
 254
 255 //
 256 // Debugging routines
 257 //
 258 void DumpScaledObject(uint64 p0, uint64 p1, uint64 p2)
 259 {
 260         WriteLog(" (SCALED BITMAP)");
 261         WriteLog(" %08X --> phrase %08X %08X\n", op_pointer, (uint32)(p1>>32), (uint32)(p1&0xFFFFFFFF));
 262         WriteLog("                 %08X --> phrase %08X %08X ", op_pointer+8, (uint32)(p2>>32), (uint32)(p2&0xFFFFFFFF));
 263         uint8 bitdepth = (p1 >> 12) & 0x07;
 264         int16 ypos = ((p0 >> 3) & 0x3FF);                       // ??? What if not interlaced (/2)?
 265         int32 xpos = p1 & 0xFFF;
 266         xpos = (xpos & 0x800 ? xpos | 0xFFFFF000 : xpos);
 267         uint32 iwidth = ((p1 >> 28) & 0x3FF);
 268         uint32 dwidth = ((p1 >> 18) & 0x3FF);           // Unsigned!
 269         uint16 height = ((p0 >> 14) & 0x3FF);
 270         uint32 link = ((p0 >> 24) & 0x7FFFF) << 3;
 271         uint32 ptr = ((p0 >> 43) & 0x1FFFFF) << 3;
 272         uint32 firstPix = (p1 >> 49) & 0x3F;
 273         uint8 flags = (p1 >> 45) & 0x0F;
 274         uint8 idx = (p1 >> 38) & 0x7F;
 275         uint32 pitch = (p1 >> 15) & 0x07;
 276         WriteLog("\n    [%u (%u) x %u @ (%i, %u) (%u bpp), l: %08X, p: %08X fp: %02X, fl:%s%s%s%s, idx:%02X, pt:%02X]\n",
 277                 iwidth, dwidth, height, xpos, ypos, op_bitmap_bit_depth[bitdepth], link, ptr, firstPix, (flags&OPFLAG_REFLECT ? "REFLECT " : ""), (flags&OPFLAG_RMW ? "RMW " : ""), (flags&OPFLAG_TRANS ? "TRANS " : ""), (flags&OPFLAG_RELEASE ? "RELEASE" : ""), idx, pitch);
 278         uint32 hscale = p2 & 0xFF;
 279         uint32 vscale = (p2 >> 8) & 0xFF;
 280         uint32 remainder = (p2 >> 16) & 0xFF;
 281         WriteLog("    [hsc: %02X, vsc: %02X, rem: %02X]\n", hscale, vscale, remainder);
 282 }
 283
 284 void DumpFixedObject(uint64 p0, uint64 p1)
 285 {
 286         WriteLog(" (BITMAP)");
 287         WriteLog(" %08X --> phrase %08X %08X\n", op_pointer, (uint32)(p1>>32), (uint32)(p1&0xFFFFFFFF));
 288         uint8 bitdepth = (p1 >> 12) & 0x07;
 289         int16 ypos = ((p0 >> 3) & 0x3FF);                       // ??? What if not interlaced (/2)?
 290         int32 xpos = p1 & 0xFFF;
 291         xpos = (xpos & 0x800 ? xpos | 0xFFFFF000 : xpos);
 292         uint32 iwidth = ((p1 >> 28) & 0x3FF);
 293         uint32 dwidth = ((p1 >> 18) & 0x3FF);           // Unsigned!
 294         uint16 height = ((p0 >> 14) & 0x3FF);
 295         uint32 link = ((p0 >> 24) & 0x7FFFF) << 3;
 296         uint32 ptr = ((p0 >> 43) & 0x1FFFFF) << 3;
 297         uint32 firstPix = (p1 >> 49) & 0x3F;
 298         uint8 flags = (p1 >> 45) & 0x0F;
 299         uint8 idx = (p1 >> 38) & 0x7F;
 300         uint32 pitch = (p1 >> 15) & 0x07;
 301         WriteLog("    [%u (%u) x %u @ (%i, %u) (%u bpp), l: %08X, p: %08X fp: %02X, fl:%s%s%s%s, idx:%02X, pt:%02X]\n",
 302                 iwidth, dwidth, height, xpos, ypos, op_bitmap_bit_depth[bitdepth], link, ptr, firstPix, (flags&OPFLAG_REFLECT ? "REFLECT " : ""), (flags&OPFLAG_RMW ? "RMW " : ""), (flags&OPFLAG_TRANS ? "TRANS " : ""), (flags&OPFLAG_RELEASE ? "RELEASE" : ""), idx, pitch);
 303 }
 304
 305 //
 306 // Object Processor main routine
 307 //
 308 void OPProcessList(int scanline, bool render)
 309 {
 310 extern int op_start_log;
 311 //      char * condition_to_str[8] =
 312 //              { "==", "<", ">", "(opflag set)", "(second half line)", "?", "?", "?" };
 313
 314         op_pointer = op_get_list_pointer();
 315
 316         objectp_stop_reading_list = false;
 317
 318 // *** BEGIN OP PROCESSOR TESTING ONLY ***
 319 extern bool interactiveMode;
 320 extern bool iToggle;
 321 extern int objectPtr;
 322 bool inhibit;
 323 int bitmapCounter = 0;
 324 // *** END OP PROCESSOR TESTING ONLY ***
 325
 326 //      if (op_pointer) WriteLog(" new op list at 0x%.8x scanline %i\n",op_pointer,scanline);
 327         while (op_pointer)
 328         {
 329 // *** BEGIN OP PROCESSOR TESTING ONLY ***
 330 if (interactiveMode && bitmapCounter == objectPtr)
 331         inhibit = iToggle;
 332 else
 333         inhibit = false;
 334 // *** END OP PROCESSOR TESTING ONLY ***
 335                 if (objectp_stop_reading_list)
 336                         return;
 337
 338                 uint64 p0 = op_load_phrase(op_pointer);
 339                 op_pointer += 8;
 340 if (scanline == tom_get_vdb() && op_start_log)
 341 //if (scanline == 215 && op_start_log)
 342 {
 343 WriteLog("%08X --> phrase %08X %08X", op_pointer - 8, (int)(p0>>32), (int)(p0&0xFFFFFFFF));
 344 if ((p0 & 0x07) == OBJECT_TYPE_BITMAP)
 345 {
 346 WriteLog(" (BITMAP) ");
 347 uint64 p1 = op_load_phrase(op_pointer);
 348 WriteLog("\n%08X --> phrase %08X %08X ", op_pointer, (int)(p1>>32), (int)(p1&0xFFFFFFFF));
 349         uint8 bitdepth = (p1 >> 12) & 0x07;
 350         int16 ypos = ((p0 >> 3) & 0x3FF);                       // ??? What if not interlaced (/2)?
 351 int32 xpos = p1 & 0xFFF;
 352 xpos = (xpos & 0x800 ? xpos | 0xFFFFF000 : xpos);
 353         uint32 iwidth = ((p1 >> 28) & 0x3FF);
 354         uint32 dwidth = ((p1 >> 18) & 0x3FF);           // Unsigned!
 355         uint16 height = ((p0 >> 14) & 0x3FF);
 356         uint32 link = ((p0 >> 24) & 0x7FFFF) << 3;
 357         uint32 ptr = ((p0 >> 43) & 0x1FFFFF) << 3;
 358         uint32 firstPix = (p1 >> 49) & 0x3F;
 359         uint8 flags = (p1 >> 45) & 0x0F;
 360         uint8 idx = (p1 >> 38) & 0x7F;
 361         uint32 pitch = (p1 >> 15) & 0x07;
 362 WriteLog("\n    [%u (%u) x %u @ (%i, %u) (%u bpp), l: %08X, p: %08X fp: %02X, fl:%s%s%s%s, idx:%02X, pt:%02X]\n",
 363         iwidth, dwidth, height, xpos, ypos, op_bitmap_bit_depth[bitdepth], link, ptr, firstPix, (flags&OPFLAG_REFLECT ? "REFLECT " : ""), (flags&OPFLAG_RMW ? "RMW " : ""), (flags&OPFLAG_TRANS ? "TRANS " : ""), (flags&OPFLAG_RELEASE ? "RELEASE" : ""), idx, pitch);
 364 }
 365 if ((p0 & 0x07) == OBJECT_TYPE_SCALE)
 366 {
 367 WriteLog(" (SCALED BITMAP)");
 368 uint64 p1 = op_load_phrase(op_pointer), p2 = op_load_phrase(op_pointer+8);
 369 WriteLog("\n%08X --> phrase %08X %08X ", op_pointer, (int)(p1>>32), (int)(p1&0xFFFFFFFF));
 370 WriteLog("\n%08X --> phrase %08X %08X ", op_pointer+8, (int)(p2>>32), (int)(p2&0xFFFFFFFF));
 371         uint8 bitdepth = (p1 >> 12) & 0x07;
 372         int16 ypos = ((p0 >> 3) & 0x3FF);                       // ??? What if not interlaced (/2)?
 373 int32 xpos = p1 & 0xFFF;
 374 xpos = (xpos & 0x800 ? xpos | 0xFFFFF000 : xpos);
 375         uint32 iwidth = ((p1 >> 28) & 0x3FF);
 376         uint32 dwidth = ((p1 >> 18) & 0x3FF);           // Unsigned!
 377         uint16 height = ((p0 >> 14) & 0x3FF);
 378         uint32 link = ((p0 >> 24) & 0x7FFFF) << 3;
 379         uint32 ptr = ((p0 >> 43) & 0x1FFFFF) << 3;
 380         uint32 firstPix = (p1 >> 49) & 0x3F;
 381         uint8 flags = (p1 >> 45) & 0x0F;
 382         uint8 idx = (p1 >> 38) & 0x7F;
 383         uint32 pitch = (p1 >> 15) & 0x07;
 384 WriteLog("\n    [%u (%u) x %u @ (%i, %u) (%u bpp), l: %08X, p: %08X fp: %02X, fl:%s%s%s%s, idx:%02X, pt:%02X]\n",
 385         iwidth, dwidth, height, xpos, ypos, op_bitmap_bit_depth[bitdepth], link, ptr, firstPix, (flags&OPFLAG_REFLECT ? "REFLECT " : ""), (flags&OPFLAG_RMW ? "RMW " : ""), (flags&OPFLAG_TRANS ? "TRANS " : ""), (flags&OPFLAG_RELEASE ? "RELEASE" : ""), idx, pitch);
 386         uint32 hscale = p2 & 0xFF;
 387         uint32 vscale = (p2 >> 8) & 0xFF;
 388         uint32 remainder = (p2 >> 16) & 0xFF;
 389 WriteLog("    [hsc: %02X, vsc: %02X, rem: %02X]\n", hscale, vscale, remainder);
 390 }
 391 if ((p0 & 0x07) == OBJECT_TYPE_GPU)
 392 WriteLog(" (GPU)\n");
 393 if ((p0 & 0x07) == OBJECT_TYPE_BRANCH)
 394 {
 395 WriteLog(" (BRANCH)\n");
 396 uint8 * jaguar_mainRam = GetRamPtr();
 397 WriteLog("[RAM] --> ");
 398 for(int k=0; k<8; k++)
 399         WriteLog("%02X ", jaguar_mainRam[op_pointer-8 + k]);
 400 WriteLog("\n");
 401 }
 402 if ((p0 & 0x07) == OBJECT_TYPE_STOP)
 403 WriteLog("    --> List end\n");
 404 }//*/
 405
 406 //              WriteLog("%08X type %i\n", op_pointer, (uint8)p0 & 0x07);
 407                 switch ((uint8)p0 & 0x07)
 408                 {
 409                 case OBJECT_TYPE_BITMAP:
 410                 {
 411                         // Would *not* be /2 if interlaced...!
 412                         uint16 ypos = ((p0 >> 3) & 0x3FF) / 2;
 413 // This is only theory implied by Rayman...!
 414 // It seems that if the YPOS is zero, then bump the YPOS value so that it coincides with
 415 // the VDB value. With interlacing, this would be slightly more tricky.
 416 // There's probably another bit somewhere that enables this mode--but so far, doesn't seem
 417 // to affect any other game in a negative way (that I've seen).
 418 // Either that, or it's an undocumented bug...
 419
 420 //No, the reason this was needed is that the OP code before was wrong. Any value
 421 //less than VDB will get written to the top line of the display!
 422 //                      if (ypos == 0)
 423 //                              ypos = TOMReadWord(0xF00046, OP) / 2;                   // Get the VDB value
 424                         uint32 height = (p0 & 0xFFC000) >> 14;
 425                         uint32 oldOPP = op_pointer - 8;
 426 // *** BEGIN OP PROCESSOR TESTING ONLY ***
 427 if (inhibit && op_start_log)
 428         WriteLog("!!! ^^^ This object is INHIBITED! ^^^ !!!\n");
 429 bitmapCounter++;
 430 if (!inhibit)   // For OP testing only!
 431 // *** END OP PROCESSOR TESTING ONLY ***
 432                         if (scanline >= ypos && height > 0)
 433                         {
 434                                 uint64 p1 = op_load_phrase(op_pointer);
 435                                 op_pointer += 8;
 436 //WriteLog("OP: Writing scanline %d with ypos == %d...\n", scanline, ypos);
 437 //WriteLog("--> Writing %u BPP bitmap...\n", op_bitmap_bit_depth[(p1 >> 12) & 0x07]);
 438                                 OPProcessFixedBitmap(scanline, p0, p1, render);
 439
 440                                 // OP write-backs
 441
 442 //???Does this really happen??? Doesn't seem to work if you do this...!
 443 //                              uint32 link = (p0 & 0x7FFFF000000) >> 21;
 444 //                              SET16(objectp_ram, 0x20, link & 0xFFFF);        // OLP
 445 //                              SET16(objectp_ram, 0x22, link >> 16);
 446 /*                              uint32 height = (p0 & 0xFFC000) >> 14;
 447                                 if (height - 1 > 0)
 448                                         height--;*/
 449                                 // NOTE: Would subtract 2 if in interlaced mode...!
 450 //                              uint64 height = ((p0 & 0xFFC000) - 0x4000) & 0xFFC000;
 451 //                              if (height)
 452                                         height--;
 453
 454                                 uint64 data = (p0 & 0xFFFFF80000000000) >> 40;
 455                                 uint64 dwidth = (p1 & 0xFFC0000) >> 15;
 456                                 data += dwidth;
 457
 458                                 p0 &= ~0xFFFFF80000FFC000;                      // Mask out old data...
 459                                 p0 |= (uint64)height << 14;
 460                                 p0 |= data << 40;
 461                                 OPStorePhrase(oldOPP, p0);
 462                         }
 463                         op_pointer = (p0 & 0x000007FFFF000000) >> 21;
 464                         break;
 465                 }
 466                 case OBJECT_TYPE_SCALE:
 467                 {
 468                         // Would *not* be /2 if interlaced...!
 469                         uint16 ypos = ((p0 >> 3) & 0x3FF) / 2;
 470 // This is only theory implied by Rayman...!
 471 // It seems that if the YPOS is zero, then bump the YPOS value so that it coincides with
 472 // the VDB value. With interlacing, this would be slightly more tricky.
 473 // There's probably another bit somewhere that enables this mode--but so far, doesn't seem
 474 // to affect any other game in a negative way (that I've seen).
 475 // Either that, or it's an undocumented bug...
 476
 477 //No, the reason this was needed is that the OP code before was wrong. Any value
 478 //less than VDB will get written to the top line of the display!
 479 //                      if (ypos == 0)
 480 //                              ypos = TOMReadWord(0xF00046, OP) / 2;                   // Get the VDB value
 481                         uint32 height = (p0 & 0xFFC000) >> 14;
 482                         uint32 oldOPP = op_pointer - 8;
 483 // *** BEGIN OP PROCESSOR TESTING ONLY ***
 484 if (inhibit && op_start_log)
 485 {
 486         WriteLog("!!! ^^^ This object is INHIBITED! ^^^ !!! (scanline=%u, ypos=%u, height=%u)\n", scanline, ypos, height);
 487         DumpScaledObject(p0, op_load_phrase(op_pointer), op_load_phrase(op_pointer+8));
 488 }
 489 bitmapCounter++;
 490 if (!inhibit)   // For OP testing only!
 491 // *** END OP PROCESSOR TESTING ONLY ***
 492                         if (scanline >= ypos && height > 0)
 493                         {
 494                                 uint64 p1 = op_load_phrase(op_pointer);
 495                                 op_pointer += 8;
 496                                 uint64 p2 = op_load_phrase(op_pointer);
 497                                 op_pointer += 8;
 498 //WriteLog("OP: %08X (%d) %08X%08X %08X%08X %08X%08X\n", oldOPP, scanline, (uint32)(p0>>32), (uint32)(p0&0xFFFFFFFF), (uint32)(p1>>32), (uint32)(p1&0xFFFFFFFF), (uint32)(p2>>32), (uint32)(p2&0xFFFFFFFF));
 499                                 OPProcessScaledBitmap(scanline, p0, p1, p2, render);
 500
 501                                 // OP write-backs
 502
 503 //???Does this really happen??? Doesn't seem to work if you do this...!
 504 //                              uint32 link = (p0 & 0x7FFFF000000) >> 21;
 505 //                              SET16(objectp_ram, 0x20, link & 0xFFFF);        // OLP
 506 //                              SET16(objectp_ram, 0x22, link >> 16);
 507 /*                              uint32 height = (p0 & 0xFFC000) >> 14;
 508                                 if (height - 1 > 0)
 509                                         height--;*/
 510                                 // NOTE: Would subtract 2 if in interlaced mode...!
 511 //                              uint64 height = ((p0 & 0xFFC000) - 0x4000) & 0xFFC000;
 512
 513                                 uint8 remainder = p2 >> 16, vscale = p2 >> 8;
 514 //Actually, we should skip this object if it has a vscale of zero.
 515 //Or do we? Not sure... Atari Karts has a few lines that look like:
 516 // (SCALED BITMAP)
 517 //000E8268 --> phrase 00010000 7000B00D
 518 //    [7 (0) x 1 @ (13, 0) (8 bpp), l: 000E82A0, p: 000E0FC0 fp: 00, fl:RELEASE, idx:00, pt:01]
 519 //    [hsc: 9A, vsc: 00, rem: 00]
 520 // Could it be the vscale is overridden if the DWIDTH is zero? Hmm...
 521
 522                                 if (vscale == 0)
 523                                         vscale = 0x20;                                  // OP bug??? Nope, it isn't...! Or is it?
 524
 525                                 remainder -= 0x20;                                      // 1.0f in [3.5] fixed point format
 526                                 if (remainder & 0x80)                           // I.e., it's negative
 527                                 {
 528                                         uint64 data = (p0 & 0xFFFFF80000000000) >> 40;
 529                                         uint64 dwidth = (p1 & 0xFFC0000) >> 15;
 530
 531                                         while (remainder & 0x80)
 532                                         {
 533                                                 remainder += vscale;
 534                                                 if (height)
 535                                                         height--;
 536
 537                                                 data += dwidth;
 538                                         }
 539                                         p0 &= ~0xFFFFF80000FFC000;              // Mask out old data...
 540                                         p0 |= (uint64)height << 14;
 541                                         p0 |= data << 40;
 542                                         OPStorePhrase(oldOPP, p0);
 543                                 }
 544
 545 //WriteLog(" [%08X%08X -> ", (uint32)(p2>>32), (uint32)(p2&0xFFFFFFFF));
 546                                 p2 &= ~0x0000000000FF0000;
 547                                 p2 |= (uint64)remainder << 16;
 548 //WriteLog("%08X%08X]\n", (uint32)(p2>>32), (uint32)(p2&0xFFFFFFFF));
 549                                 OPStorePhrase(oldOPP+16, p2);
 550 //remainder = (uint8)(p2 >> 16), vscale = (uint8)(p2 >> 8);
 551 //WriteLog(" [after]: rem=%02X, vscale=%02X\n", remainder, vscale);
 552                         }
 553                         op_pointer = (p0 & 0x000007FFFF000000) >> 21;
 554                         break;
 555                 }
 556                 case OBJECT_TYPE_GPU:
 557                 {
 558 //WriteLog("OP: Asserting GPU IRQ #3...\n");
 559                         op_set_current_object(p0);
 560                         GPUSetIRQLine(3, ASSERT_LINE);
 561 //Also, OP processing is suspended from this point until OBF (F00026) is written to...
 562 // !!! FIX !!!
 563 //Do something like:
 564 //OPSuspendedByGPU = true;
 565 //Dunno if the OP keeps processing from where it was interrupted, or if it just continues
 566 //on the next scanline...
 567                         break;
 568                 }
 569                 case OBJECT_TYPE_BRANCH:
 570                 {
 571                         uint16 ypos = (p0 >> 3) & 0x7FF;
 572                         uint8  cc   = (p0 >> 14) & 0x03;
 573                         uint32 link = (p0 >> 21) & 0x3FFFF8;
 574
 575 //                      if ((ypos!=507)&&(ypos!=25))
 576 //                              WriteLog("\t%i%s%i link=0x%.8x\n",scanline,condition_to_str[cc],ypos>>1,link);
 577                         switch (cc)
 578                         {
 579                         case CONDITION_EQUAL:
 580 //Why do this for the equal case? If they wrote an odd YPOS, then it wouldn't be detected!
 581 //                              if (ypos != 0x7FF && (ypos & 0x01))
 582 //                                      ypos ^= 0x01;
 583 //                              if ((2 * tom_get_scanline()) == ypos || ypos == 0x7FF)
 584 //Here we're using VC instead of the bogus tom_get_scanline() value...
 585                                 if (TOMReadWord(0xF00006, OP) == ypos || ypos == 0x7FF)
 586                                         op_pointer = link;
 587                                 break;
 588                         case CONDITION_LESS_THAN:
 589 //                              if ((2 * tom_get_scanline()) < ypos)
 590                                 if (TOMReadWord(0xF00006, OP) < ypos)
 591                                         op_pointer = link;
 592                                 break;
 593                         case CONDITION_GREATER_THAN:
 594 //                              if ((2 * tom_get_scanline()) > ypos)
 595                                 if (TOMReadWord(0xF00006, OP) > ypos)
 596                                         op_pointer = link;
 597                                 break;
 598                         case CONDITION_OP_FLAG_SET:
 599                                 if (op_get_status_register() & 0x01)
 600                                         op_pointer = link;
 601                                 break;
 602                         case CONDITION_SECOND_HALF_LINE:
 603                                 // This basically means branch if bit 10 of HC is set
 604                                 WriteLog("OP: Unexpected CONDITION_SECOND_HALF_LINE in BRANCH object\nOP: shuting down\n");
 605                                 fclose(log_get());
 606                                 exit(0);
 607                                 break;
 608                         default:
 609                                 WriteLog("OP: Unimplemented branch condition %i\n", cc);
 610                         }
 611                         break;
 612                 }
 613                 case OBJECT_TYPE_STOP:
 614                 {
 615 //op_start_log = 0;
 616                         // unsure
 617 //WriteLog("OP: --> STOP\n");
 618 //                      op_set_status_register(((p0>>3) & 0xFFFFFFFF));
 619 //This seems more likely...
 620                         op_set_current_object(p0);
 621
 622                         if (p0 & 0x08)
 623                         {
 624                                 tom_set_pending_object_int();
 625                                 if (tom_irq_enabled(IRQ_OPFLAG) && jaguar_interrupt_handler_is_valid(64))
 626                                         m68k_set_irq(7);                                // Cause an NMI to occur...
 627                         }
 628
 629                         return;
 630 //                      break;
 631                 }
 632                 default:
 633                         WriteLog("op: unknown object type %i\n", ((uint8)p0 & 0x07));
 634                         return;
 635                 }
 636         }
 637 }
 638
 639 //
 640 // Store fixed size bitmap in line buffer
 641 //
 642
 643 // Interesting thing about Rayman: There seems to be a transparent bitmap (1/8/16 bpp--which?)
 644 // being rendered under his feet--doesn't align when walking... Check it out!
 645
 646 void OPProcessFixedBitmap(int scanline, uint64 p0, uint64 p1, bool render)
 647 {
 648 // Need to make sure that when writing that it stays within the line buffer...
 649 // LBUF ($F01800 - $F01D9E) 360 x 32-bit RAM
 650         uint8 depth = (p1 >> 12) & 0x07;                                // Color depth of image
 651         int32 xpos = ((int16)((p1 << 4) & 0xFFFF)) >> 4;// Image xpos in LBUF
 652         uint32 iwidth = (p1 >> 28) & 0x3FF;                             // Image width in *phrases*
 653         uint32 data = (p0 >> 40) & 0xFFFFF8;                    // Pixel data address
 654 //#ifdef OP_DEBUG_BMP
 655 // Prolly should use this... Though not sure exactly how.
 656         uint32  firstPix = (p1 >> 49) & 0x3F;
 657         // "The LSB is significant only for scaled objects..." -JTRM
 658         // "In 1 BPP mode, all five bits are significant. In 2 BPP mode, the top four are significant..."
 659         firstPix &= 0x3E;
 660 //#endif
 661 // We can ignore the RELEASE (high order) bit for now--probably forever...!
 662 //      uint8 flags = (p1 >> 45) & 0x0F;        // REFLECT, RMW, TRANS, RELEASE
 663 //Optimize: break these out to their own BOOL values
 664         uint8 flags = (p1 >> 45) & 0x07;                                // REFLECT (0), RMW (1), TRANS (2)
 665         bool flagREFLECT = (flags & OPFLAG_REFLECT ? true : false),
 666                 flagRMW = (flags & OPFLAG_RMW ? true : false),
 667                 flagTRANS = (flags & OPFLAG_TRANS ? true : false);
 668 // "For images with 1 to 4 bits/pixel the top 7 to 4 bits of the index
 669 //  provide the most significant bits of the palette address."
 670         uint8 index = (p1 >> 37) & 0xFE;                                // CLUT index offset (upper pix, 1-4 bpp)
 671         uint32 pitch = (p1 >> 15) & 0x07;                               // Phrase pitch
 672         pitch <<= 3;                                                                    // Optimization: Multiply pitch by 8
 673
 674 //      int16 scanlineWidth = tom_getVideoModeWidth();
 675         uint8 * tom_ram_8 = tom_get_ram_pointer();
 676         uint8 * paletteRAM = &tom_ram_8[0x400];
 677         // This is OK as long as it's used correctly: For 16-bit RAM to RAM direct copies--NOT
 678         // for use when using endian-corrected data (i.e., any of the *_word_read functions!)
 679         uint16 * paletteRAM16 = (uint16 *)paletteRAM;
 680
 681 //      WriteLog("bitmap %ix? %ibpp at %i,? firstpix=? data=0x%.8x pitch %i hflipped=%s dwidth=? (linked to ?) RMW=%s Tranparent=%s\n",
 682 //              iwidth, op_bitmap_bit_depth[bitdepth], xpos, ptr, pitch, (flags&OPFLAG_REFLECT ? "yes" : "no"), (flags&OPFLAG_RMW ? "yes" : "no"), (flags&OPFLAG_TRANS ? "yes" : "no"));
 683
 684 // Is it OK to have a 0 for the data width??? (i.e., undocumented?)
 685 // Seems to be... Seems that dwidth *can* be zero (i.e., reuse same line) as well.
 686 // Pitch == 0 is OK too...
 687 //      if (!render || op_pointer == 0 || ptr == 0 || pitch == 0)
 688 //I'm not convinced that we need to concern ourselves with data & op_pointer here either!
 689         if (!render || iwidth == 0)
 690                 return;
 691
 692 //#define OP_DEBUG_BMP
 693 //#ifdef OP_DEBUG_BMP
 694 //      WriteLog("bitmap %ix%i %ibpp at %i,%i firstpix=%i data=0x%.8x pitch %i hflipped=%s dwidth=%i (linked to 0x%.8x) Transluency=%s\n",
 695 //              iwidth, height, op_bitmap_bit_depth[bitdepth], xpos, ypos, firstPix, ptr, pitch, (flags&OPFLAG_REFLECT ? "yes" : "no"), dwidth, op_pointer, (flags&OPFLAG_RMW ? "yes" : "no"));
 696 //#endif
 697
 698 //      int32 leftMargin = xpos, rightMargin = (xpos + (phraseWidthToPixels[depth] * iwidth)) - 1;
 699         int32 startPos = xpos, endPos = xpos +
 700                 (!flagREFLECT ? (phraseWidthToPixels[depth] * iwidth) - 1
 701                 : -((phraseWidthToPixels[depth] * iwidth) + 1));
 702         uint32 clippedWidth = 0, phraseClippedWidth = 0, dataClippedWidth = 0;//, phrasePixel = 0;
 703         bool in24BPPMode = (((GET16(tom_ram_8, 0x0028) >> 1) & 0x03) == 1 ? true : false);      // VMODE
 704         // Not sure if this is Jaguar Two only location or what...
 705         // From the docs, it is... If we want to limit here we should think of something else.
 706 //      int32 limit = GET16(tom_ram_8, 0x0008);                 // LIMIT
 707         int32 limit = 720;
 708         int32 lbufWidth = (!in24BPPMode ? limit - 1 : (limit / 2) - 1); // Zero based limit...
 709
 710         // If the image is completely to the left or right of the line buffer, then bail.
 711 //If in REFLECT mode, then these values are swapped! !!! FIX !!! [DONE]
 712 //There are four possibilities:
 713 //  1. image sits on left edge and no REFLECT; starts out of bounds but ends in bounds.
 714 //  2. image sits on left edge and REFLECT; starts in bounds but ends out of bounds.
 715 //  3. image sits on right edge and REFLECT; starts out of bounds but ends in bounds.
 716 //  4. image sits on right edge and no REFLECT; starts in bounds but ends out of bounds.
 717 //Numbers 2 & 4 can be caught by checking the LBUF clip while in the inner loop,
 718 // numbers 1 & 3 are of concern.
 719 // This *indirectly* handles only cases 2 & 4! And is WRONG is REFLECT is set...!
 720 //      if (rightMargin < 0 || leftMargin > lbufWidth)
 721
 722 // It might be easier to swap these (if REFLECTed) and just use XPOS down below...
 723 // That way, you could simply set XPOS to leftMargin if !REFLECT and to rightMargin otherwise.
 724 // Still have to be careful with the DATA and IWIDTH values though...
 725
 726 //      if ((!flagREFLECT && (rightMargin < 0 || leftMargin > lbufWidth))
 727 //              || (flagREFLECT && (leftMargin < 0 || rightMargin > lbufWidth)))
 728 //              return;
 729         if ((!flagREFLECT && (endPos < 0 || startPos > lbufWidth))
 730                 || (flagREFLECT && (startPos < 0 || endPos > lbufWidth)))
 731                 return;
 732
 733         // Otherwise, find the clip limits and clip the phrase as well...
 734         // NOTE: I'm fudging here by letting the actual blit overstep the bounds of the
 735         //       line buffer, but it shouldn't matter since there are two unused line
 736         //       buffers below and nothing above and I'll at most write 8 bytes outside
 737         //       the line buffer... I could use a fractional clip begin/end value, but
 738         //       this makes the blit a *lot* more hairy. I might fix this in the future
 739         //       if it becomes necessary. (JLH)
 740         //       Probably wouldn't be *that* hairy. Just use a delta that tells the inner loop
 741         //       which pixel in the phrase is being written, and quit when either end of phrases
 742         //       is reached or line buffer extents are surpassed.
 743
 744 //This stuff is probably wrong as well... !!! FIX !!!
 745 //The strange thing is that it seems to work, but that's no guarantee that it's bulletproof!
 746 //Yup. Seems that JagMania doesn't work correctly with this...
 747 //Dunno if this is the problem, but Atari Karts is showing *some* of the road now...
 748 //      if (!flagREFLECT)
 749
 750 /*
 751         if (leftMargin < 0)
 752                 clippedWidth = 0 - leftMargin,
 753                 phraseClippedWidth = clippedWidth / phraseWidthToPixels[depth],
 754                 leftMargin = 0 - (clippedWidth % phraseWidthToPixels[depth]);
 755 //              leftMargin = 0;
 756
 757         if (rightMargin > lbufWidth)
 758                 clippedWidth = rightMargin - lbufWidth,
 759                 phraseClippedWidth = clippedWidth / phraseWidthToPixels[depth];//,
 760 //              rightMargin = lbufWidth + (clippedWidth % phraseWidthToPixels[depth]);
 761 //              rightMargin = lbufWidth;
 762 */
 763 if (depth > 5)
 764         WriteLog("We're about to encounter a divide by zero error!\n");
 765         // NOTE: We're just using endPos to figure out how much, if any, to clip by.
 766         // ALSO: There may be another case where we start out of bounds and end out of bounds...!
 767         // !!! FIX !!!
 768         if (startPos < 0)                       // Case #1: Begin out, end in, L to R
 769                 clippedWidth = 0 - startPos,
 770                 dataClippedWidth = phraseClippedWidth = clippedWidth / phraseWidthToPixels[depth],
 771                 startPos = 0 - (clippedWidth % phraseWidthToPixels[depth]);
 772
 773         if (endPos < 0)                         // Case #2: Begin in, end out, R to L
 774                 clippedWidth = 0 - endPos,
 775                 phraseClippedWidth = clippedWidth / phraseWidthToPixels[depth];
 776
 777         if (endPos > lbufWidth)         // Case #3: Begin in, end out, L to R
 778                 clippedWidth = endPos - lbufWidth,
 779                 phraseClippedWidth = clippedWidth / phraseWidthToPixels[depth];
 780
 781         if (startPos > lbufWidth)       // Case #4: Begin out, end in, R to L
 782                 clippedWidth = startPos - lbufWidth,
 783                 dataClippedWidth = phraseClippedWidth = clippedWidth / phraseWidthToPixels[depth],
 784                 startPos = lbufWidth + (clippedWidth % phraseWidthToPixels[depth]);
 785
 786         // If the image is sitting on the line buffer left or right edge, we need to compensate
 787         // by decreasing the image phrase width accordingly.
 788         iwidth -= phraseClippedWidth;
 789
 790         // Also, if we're clipping the phrase we need to make sure we're in the correct part of
 791         // the pixel data.
 792 //      data += phraseClippedWidth * (pitch << 3);
 793         data += dataClippedWidth * pitch;
 794
 795         // NOTE: When the bitmap is in REFLECT mode, the XPOS marks the *right* side of the
 796         //       bitmap! This makes clipping & etc. MUCH, much easier...!
 797 //      uint32 lbufAddress = 0x1800 + (!in24BPPMode ? leftMargin * 2 : leftMargin * 4);
 798 //Why does this work right when multiplying startPos by 2 (instead of 4) for 24 BPP mode?
 799 //Is this a bug in the OP?
 800         uint32 lbufAddress = 0x1800 + (!in24BPPMode ? startPos * 2 : startPos * 2);
 801         uint8 * currentLineBuffer = &tom_ram_8[lbufAddress];
 802
 803         // Render.
 804
 805 // Hmm. We check above for 24 BPP mode, but don't do anything about it below...
 806 // If we *were* in 24 BPP mode, how would you convert CRY to RGB24? Seems to me
 807 // that if you're in CRY mode then you wouldn't be able to use 24 BPP bitmaps
 808 // anyway.
 809 // This seems to be the case (at least according to the Midsummer docs)...!
 810
 811         if (depth == 0)                                                                 // 1 BPP
 812         {
 813                 // The LSB of flags is OPFLAG_REFLECT, so sign extend it and or 2 into it.
 814                 int32 lbufDelta = ((int8)((flags << 7) & 0xFF) >> 5) | 0x02;
 815
 816                 // Fetch 1st phrase...
 817                 uint64 pixels = ((uint64)JaguarReadLong(data, OP) << 32) | JaguarReadLong(data + 4, OP);
 818 //Note that firstPix should only be honored *if* we start with the 1st phrase of the bitmap
 819 //i.e., we didn't clip on the margin... !!! FIX !!!
 820                 pixels <<= firstPix;                                            // Skip first N pixels (N=firstPix)...
 821                 int i = firstPix;                                                       // Start counter at right spot...
 822
 823                 while (iwidth--)
 824                 {
 825                         while (i++ < 64)
 826                         {
 827                                 uint8 bit = pixels >> 63;
 828                                 if (flagTRANS && bit == 0)
 829                                         ;       // Do nothing...
 830                                 else
 831                                 {
 832                                         if (!flagRMW)
 833 //Optimize: Set palleteRAM16 to beginning of palette RAM + index*2 and use only [bit] as index...
 834 //Won't optimize RMW case though...
 835                                                 // This is the *only* correct use of endian-dependent code
 836                                                 // (i.e., mem-to-mem direct copying)!
 837                                                 *(uint16 *)currentLineBuffer = paletteRAM16[index | bit];
 838                                         else
 839                                                 *currentLineBuffer =
 840                                                         BLEND_CR(*currentLineBuffer, paletteRAM[(index | bit) << 1]),
 841                                                 *(currentLineBuffer + 1) =
 842                                                         BLEND_Y(*(currentLineBuffer + 1), paletteRAM[((index | bit) << 1) + 1]);
 843                                 }
 844
 845                                 currentLineBuffer += lbufDelta;
 846                                 pixels <<= 1;
 847                         }
 848                         i = 0;
 849                         // Fetch next phrase...
 850                         data += pitch;
 851                         pixels = ((uint64)JaguarReadLong(data, OP) << 32) | JaguarReadLong(data + 4, OP);
 852                 }
 853         }
 854         else if (depth == 1)                                                    // 2 BPP
 855         {
 856 if (firstPix)
 857         WriteLog("OP: Fixed bitmap @ 2 BPP requesting FIRSTPIX! (fp=%u)\n", firstPix);
 858                 index &= 0xFC;                                                          // Top six bits form CLUT index
 859                 // The LSB is OPFLAG_REFLECT, so sign extend it and or 2 into it.
 860                 int32 lbufDelta = ((int8)((flags << 7) & 0xFF) >> 5) | 0x02;
 861
 862                 while (iwidth--)
 863                 {
 864                         // Fetch phrase...
 865                         uint64 pixels = ((uint64)JaguarReadLong(data, OP) << 32) | JaguarReadLong(data + 4, OP);
 866                         data += pitch;
 867
 868                         for(int i=0; i<32; i++)
 869                         {
 870                                 uint8 bits = pixels >> 62;
 871 // Seems to me that both of these are in the same endian, so we could cast it as
 872 // uint16 * and do straight across copies (what about 24 bpp? Treat it differently...)
 873 // This only works for the palettized modes (1 - 8 BPP), since we actually have to
 874 // copy data from memory in 16 BPP mode (or does it? Isn't this the same as the CLUT case?)
 875 // No, it isn't because we read the memory in an endian safe way--this *won't* work...
 876                                 if (flagTRANS && bits == 0)
 877                                         ;       // Do nothing...
 878                                 else
 879                                 {
 880                                         if (!flagRMW)
 881                                                 *(uint16 *)currentLineBuffer = paletteRAM16[index | bits];
 882                                         else
 883                                                 *currentLineBuffer =
 884                                                         BLEND_CR(*currentLineBuffer, paletteRAM[(index | bits) << 1]),
 885                                                 *(currentLineBuffer + 1) =
 886                                                         BLEND_Y(*(currentLineBuffer + 1), paletteRAM[((index | bits) << 1) + 1]);
 887                                 }
 888
 889                                 currentLineBuffer += lbufDelta;
 890                                 pixels <<= 2;
 891                         }
 892                 }
 893         }
 894         else if (depth == 2)                                                    // 4 BPP
 895         {
 896 if (firstPix)
 897         WriteLog("OP: Fixed bitmap @ 4 BPP requesting FIRSTPIX! (fp=%u)\n", firstPix);
 898                 index &= 0xF0;                                                          // Top four bits form CLUT index
 899                 // The LSB is OPFLAG_REFLECT, so sign extend it and or 2 into it.
 900                 int32 lbufDelta = ((int8)((flags << 7) & 0xFF) >> 5) | 0x02;
 901
 902                 while (iwidth--)
 903                 {
 904                         // Fetch phrase...
 905                         uint64 pixels = ((uint64)JaguarReadLong(data, OP) << 32) | JaguarReadLong(data + 4, OP);
 906                         data += pitch;
 907
 908                         for(int i=0; i<16; i++)
 909                         {
 910                                 uint8 bits = pixels >> 60;
 911 // Seems to me that both of these are in the same endian, so we could cast it as
 912 // uint16 * and do straight across copies (what about 24 bpp? Treat it differently...)
 913 // This only works for the palettized modes (1 - 8 BPP), since we actually have to
 914 // copy data from memory in 16 BPP mode (or does it? Isn't this the same as the CLUT case?)
 915 // No, it isn't because we read the memory in an endian safe way--this *won't* work...
 916                                 if (flagTRANS && bits == 0)
 917                                         ;       // Do nothing...
 918                                 else
 919                                 {
 920                                         if (!flagRMW)
 921                                                 *(uint16 *)currentLineBuffer = paletteRAM16[index | bits];
 922                                         else
 923                                                 *currentLineBuffer =
 924                                                         BLEND_CR(*currentLineBuffer, paletteRAM[(index | bits) << 1]),
 925                                                 *(currentLineBuffer + 1) =
 926                                                         BLEND_Y(*(currentLineBuffer + 1), paletteRAM[((index | bits) << 1) + 1]);
 927                                 }
 928
 929                                 currentLineBuffer += lbufDelta;
 930                                 pixels <<= 4;
 931                         }
 932                 }
 933         }
 934         else if (depth == 3)                                                    // 8 BPP
 935         {
 936 if (firstPix)
 937         WriteLog("OP: Fixed bitmap @ 8 BPP requesting FIRSTPIX! (fp=%u)\n", firstPix);
 938                 // The LSB is OPFLAG_REFLECT, so sign extend it and or 2 into it.
 939                 int32 lbufDelta = ((int8)((flags << 7) & 0xFF) >> 5) | 0x02;
 940
 941                 while (iwidth--)
 942                 {
 943                         // Fetch phrase...
 944                         uint64 pixels = ((uint64)JaguarReadLong(data, OP) << 32) | JaguarReadLong(data + 4, OP);
 945                         data += pitch;
 946
 947                         for(int i=0; i<8; i++)
 948                         {
 949                                 uint8 bits = pixels >> 56;
 950 // Seems to me that both of these are in the same endian, so we could cast it as
 951 // uint16 * and do straight across copies (what about 24 bpp? Treat it differently...)
 952 // This only works for the palettized modes (1 - 8 BPP), since we actually have to
 953 // copy data from memory in 16 BPP mode (or does it? Isn't this the same as the CLUT case?)
 954 // No, it isn't because we read the memory in an endian safe way--this *won't* work...
 955                                 if (flagTRANS && bits == 0)
 956                                         ;       // Do nothing...
 957                                 else
 958                                 {
 959                                         if (!flagRMW)
 960                                                 *(uint16 *)currentLineBuffer = paletteRAM16[bits];
 961                                         else
 962                                                 *currentLineBuffer =
 963                                                         BLEND_CR(*currentLineBuffer, paletteRAM[bits << 1]),
 964                                                 *(currentLineBuffer + 1) =
 965                                                         BLEND_Y(*(currentLineBuffer + 1), paletteRAM[(bits << 1) + 1]);
 966                                 }
 967
 968                                 currentLineBuffer += lbufDelta;
 969                                 pixels <<= 8;
 970                         }
 971                 }
 972         }
 973         else if (depth == 4)                                                    // 16 BPP
 974         {
 975 if (firstPix)
 976         WriteLog("OP: Fixed bitmap @ 16 BPP requesting FIRSTPIX! (fp=%u)\n", firstPix);
 977                 // The LSB is OPFLAG_REFLECT, so sign extend it and or 2 into it.
 978                 int32 lbufDelta = ((int8)((flags << 7) & 0xFF) >> 5) | 0x02;
 979
 980                 while (iwidth--)
 981                 {
 982                         // Fetch phrase...
 983                         uint64 pixels = ((uint64)JaguarReadLong(data, OP) << 32) | JaguarReadLong(data + 4, OP);
 984                         data += pitch;
 985
 986                         for(int i=0; i<4; i++)
 987                         {
 988                                 uint8 bitsHi = pixels >> 56, bitsLo = pixels >> 48;
 989 // Seems to me that both of these are in the same endian, so we could cast it as
 990 // uint16 * and do straight across copies (what about 24 bpp? Treat it differently...)
 991 // This only works for the palettized modes (1 - 8 BPP), since we actually have to
 992 // copy data from memory in 16 BPP mode (or does it? Isn't this the same as the CLUT case?)
 993 // No, it isn't because we read the memory in an endian safe way--it *won't* work...
 994                                 if (flagTRANS && (bitsLo | bitsHi) == 0)
 995                                         ;       // Do nothing...
 996                                 else
 997                                 {
 998                                         if (!flagRMW)
 999                                                 *currentLineBuffer = bitsHi,
1000                                                 *(currentLineBuffer + 1) = bitsLo;
1001                                         else
1002                                                 *currentLineBuffer =
1003                                                         BLEND_CR(*currentLineBuffer, bitsHi),
1004                                                 *(currentLineBuffer + 1) =
1005                                                         BLEND_Y(*(currentLineBuffer + 1), bitsLo);
1006                                 }
1007
1008                                 currentLineBuffer += lbufDelta;
1009                                 pixels <<= 16;
1010                         }
1011                 }
1012         }
1013         else if (depth == 5)                                                    // 24 BPP
1014         {
1015 //Looks like Iron Soldier is the only game that uses 24BPP mode...
1016 //There *might* be others...
1017 //WriteLog("OP: Writing 24 BPP bitmap!\n");
1018 if (firstPix)
1019         WriteLog("OP: Fixed bitmap @ 24 BPP requesting FIRSTPIX! (fp=%u)\n", firstPix);
1020                 // Not sure, but I think RMW only works with 16 BPP and below, and only in CRY mode...
1021                 // The LSB of flags is OPFLAG_REFLECT, so sign extend it and OR 4 into it.
1022                 int32 lbufDelta = ((int8)((flags << 7) & 0xFF) >> 4) | 0x04;
1023
1024                 while (iwidth--)
1025                 {
1026                         // Fetch phrase...
1027                         uint64 pixels = ((uint64)JaguarReadLong(data, OP) << 32) | JaguarReadLong(data + 4, OP);
1028                         data += pitch;
1029
1030                         for(int i=0; i<2; i++)
1031                         {
1032                                 // We don't use a 32-bit var here because of endian issues...!
1033                                 uint8 bits3 = pixels >> 56, bits2 = pixels >> 48,
1034                                         bits1 = pixels >> 40, bits0 = pixels >> 32;
1035
1036                                 if (flagTRANS && (bits3 | bits2 | bits1 | bits0) == 0)
1037                                         ;       // Do nothing...
1038                                 else
1039                                         *currentLineBuffer = bits3,
1040                                         *(currentLineBuffer + 1) = bits2,
1041                                         *(currentLineBuffer + 2) = bits1,
1042                                         *(currentLineBuffer + 3) = bits0;
1043
1044                                 currentLineBuffer += lbufDelta;
1045                                 pixels <<= 32;
1046                         }
1047                 }
1048         }
1049 }
1050
1051 //
1052 // Store scaled bitmap in line buffer
1053 //
1054 void OPProcessScaledBitmap(int scanline, uint64 p0, uint64 p1, uint64 p2, bool render)
1055 {
1056 // Need to make sure that when writing that it stays within the line buffer...
1057 // LBUF ($F01800 - $F01D9E) 360 x 32-bit RAM
1058         uint8 depth = (p1 >> 12) & 0x07;                                // Color depth of image
1059         int32 xpos = ((int16)((p1 << 4) & 0xFFFF)) >> 4;// Image xpos in LBUF
1060         uint32 iwidth = (p1 >> 28) & 0x3FF;                             // Image width in *phrases*
1061         uint32 data = (p0 >> 40) & 0xFFFFF8;                    // Pixel data address
1062 //#ifdef OP_DEBUG_BMP
1063 // Prolly should use this... Though not sure exactly how.
1064 //Use the upper bits as an offset into the phrase depending on the BPP. That's how!
1065         uint32 firstPix = (p1 >> 49) & 0x3F;
1066 //This is WEIRD! I'm sure I saw Atari Karts request 8 BPP FIRSTPIX! What happened???
1067 if (firstPix)
1068         WriteLog("OP: FIRSTPIX != 0! (Scaled BM)\n");
1069 //#endif
1070 // We can ignore the RELEASE (high order) bit for now--probably forever...!
1071 //      uint8 flags = (p1 >> 45) & 0x0F;        // REFLECT, RMW, TRANS, RELEASE
1072 //Optimize: break these out to their own BOOL values
1073         uint8 flags = (p1 >> 45) & 0x07;                                // REFLECT (0), RMW (1), TRANS (2)
1074         bool flagREFLECT = (flags & OPFLAG_REFLECT ? true : false),
1075                 flagRMW = (flags & OPFLAG_RMW ? true : false),
1076                 flagTRANS = (flags & OPFLAG_TRANS ? true : false);
1077         uint8 index = (p1 >> 37) & 0xFE;                                // CLUT index offset (upper pix, 1-4 bpp)
1078         uint32 pitch = (p1 >> 15) & 0x07;                               // Phrase pitch
1079
1080 //      int16 scanlineWidth = tom_getVideoModeWidth();
1081         uint8 * tom_ram_8 = tom_get_ram_pointer();
1082         uint8 * paletteRAM = &tom_ram_8[0x400];
1083         // This is OK as long as it's used correctly: For 16-bit RAM to RAM direct copies--NOT
1084         // for use when using endian-corrected data (i.e., any of the *_word_read functions!)
1085         uint16 * paletteRAM16 = (uint16 *)paletteRAM;
1086
1087         uint8 hscale = p2 & 0xFF;
1088         uint8 horizontalRemainder = hscale;                             // Not sure if it starts full, but seems reasonable
1089         int32 scaledWidthInPixels = (iwidth * phraseWidthToPixels[depth] * hscale) >> 5;
1090         uint32 scaledPhrasePixels = (phraseWidthToPixels[depth] * hscale) >> 5;
1091
1092 //      WriteLog("bitmap %ix? %ibpp at %i,? firstpix=? data=0x%.8x pitch %i hflipped=%s dwidth=? (linked to ?) RMW=%s Tranparent=%s\n",
1093 //              iwidth, op_bitmap_bit_depth[bitdepth], xpos, ptr, pitch, (flags&OPFLAG_REFLECT ? "yes" : "no"), (flags&OPFLAG_RMW ? "yes" : "no"), (flags&OPFLAG_TRANS ? "yes" : "no"));
1094
1095 //Looks like an hscale of zero means don't draw!
1096         if (!render || iwidth == 0 || hscale == 0)
1097                 return;
1098
1099 //#define OP_DEBUG_BMP
1100 //#ifdef OP_DEBUG_BMP
1101 //      WriteLog("bitmap %ix%i %ibpp at %i,%i firstpix=%i data=0x%.8x pitch %i hflipped=%s dwidth=%i (linked to 0x%.8x) Transluency=%s\n",
1102 //              iwidth, height, op_bitmap_bit_depth[bitdepth], xpos, ypos, firstPix, ptr, pitch, (flags&OPFLAG_REFLECT ? "yes" : "no"), dwidth, op_pointer, (flags&OPFLAG_RMW ? "yes" : "no"));
1103 //#endif
1104
1105         int32 startPos = xpos, endPos = xpos +
1106                 (!flagREFLECT ? scaledWidthInPixels - 1 : -(scaledWidthInPixels + 1));
1107         uint32 clippedWidth = 0, phraseClippedWidth = 0, dataClippedWidth = 0;
1108         bool in24BPPMode = (((GET16(tom_ram_8, 0x0028) >> 1) & 0x03) == 1 ? true : false);      // VMODE
1109         // Not sure if this is Jaguar Two only location or what...
1110         // From the docs, it is... If we want to limit here we should think of something else.
1111 //      int32 limit = GET16(tom_ram_8, 0x0008);                 // LIMIT
1112         int32 limit = 720;
1113         int32 lbufWidth = (!in24BPPMode ? limit - 1 : (limit / 2) - 1); // Zero based limit...
1114
1115         // If the image is completely to the left or right of the line buffer, then bail.
1116 //If in REFLECT mode, then these values are swapped! !!! FIX !!! [DONE]
1117 //There are four possibilities:
1118 //  1. image sits on left edge and no REFLECT; starts out of bounds but ends in bounds.
1119 //  2. image sits on left edge and REFLECT; starts in bounds but ends out of bounds.
1120 //  3. image sits on right edge and REFLECT; starts out of bounds but ends in bounds.
1121 //  4. image sits on right edge and no REFLECT; starts in bounds but ends out of bounds.
1122 //Numbers 2 & 4 can be caught by checking the LBUF clip while in the inner loop,
1123 // numbers 1 & 3 are of concern.
1124 // This *indirectly* handles only cases 2 & 4! And is WRONG is REFLECT is set...!
1125 //      if (rightMargin < 0 || leftMargin > lbufWidth)
1126
1127 // It might be easier to swap these (if REFLECTed) and just use XPOS down below...
1128 // That way, you could simply set XPOS to leftMargin if !REFLECT and to rightMargin otherwise.
1129 // Still have to be careful with the DATA and IWIDTH values though...
1130
1131         if ((!flagREFLECT && (endPos < 0 || startPos > lbufWidth))
1132                 || (flagREFLECT && (startPos < 0 || endPos > lbufWidth)))
1133                 return;
1134
1135         // Otherwise, find the clip limits and clip the phrase as well...
1136         // NOTE: I'm fudging here by letting the actual blit overstep the bounds of the
1137         //       line buffer, but it shouldn't matter since there are two unused line
1138         //       buffers below and nothing above and I'll at most write 40 bytes outside
1139         //       the line buffer... I could use a fractional clip begin/end value, but
1140         //       this makes the blit a *lot* more hairy. I might fix this in the future
1141         //       if it becomes necessary. (JLH)
1142         //       Probably wouldn't be *that* hairy. Just use a delta that tells the inner loop
1143         //       which pixel in the phrase is being written, and quit when either end of phrases
1144         //       is reached or line buffer extents are surpassed.
1145
1146 //This stuff is probably wrong as well... !!! FIX !!!
1147 //The strange thing is that it seems to work, but that's no guarantee that it's bulletproof!
1148 //Yup. Seems that JagMania doesn't work correctly with this...
1149 //Dunno if this is the problem, but Atari Karts is showing *some* of the road now...
1150 //Actually, it is! Or, it was. It doesn't seem to be clipping here, so the problem lies
1151 //elsewhere! Hmm. Putting the scaling code into the 1/2/8 BPP cases seems to draw the ground
1152 // a bit more accurately... Strange!
1153 //It's probably a case of the REFLECT flag being set and the background being written
1154 //from the right side of the screen...
1155 //But no, it isn't... At least if the diagnostics are telling the truth!
1156
1157         // NOTE: We're just using endPos to figure out how much, if any, to clip by.
1158         // ALSO: There may be another case where we start out of bounds and end out of bounds...!
1159         // !!! FIX !!!
1160
1161 //There's a problem here with scaledPhrasePixels in that it can be forced to zero when
1162 //the scaling factor is small. So fix it already! !!! FIX !!!
1163 /*if (scaledPhrasePixels == 0)
1164 {
1165         WriteLog("OP: [Scaled] We're about to encounter a divide by zero error!\n");
1166         DumpScaledObject(p0, p1, p2);
1167 }//*/
1168 //NOTE: I'm almost 100% sure that this is wrong... And it is! :-p
1169         if (startPos < 0)                       // Case #1: Begin out, end in, L to R
1170 /*              clippedWidth = 0 - startPos,
1171                 dataClippedWidth = phraseClippedWidth = clippedWidth / phraseWidthToPixels[depth],
1172                 startPos = 0 - (clippedWidth % phraseWidthToPixels[depth]);*/
1173                 clippedWidth = 0 - startPos,
1174                 dataClippedWidth = phraseClippedWidth = clippedWidth / scaledPhrasePixels,
1175                 startPos = 0 - (clippedWidth % scaledPhrasePixels);
1176
1177         if (endPos < 0)                         // Case #2: Begin in, end out, R to L
1178 /*              clippedWidth = 0 - endPos,
1179                 phraseClippedWidth = clippedWidth / phraseWidthToPixels[depth];*/
1180                 clippedWidth = 0 - endPos,
1181                 phraseClippedWidth = clippedWidth / scaledPhrasePixels;
1182
1183         if (endPos > lbufWidth)         // Case #3: Begin in, end out, L to R
1184 /*              clippedWidth = endPos - lbufWidth,
1185                 phraseClippedWidth = clippedWidth / phraseWidthToPixels[depth];*/
1186                 clippedWidth = endPos - lbufWidth,
1187                 phraseClippedWidth = clippedWidth / scaledPhrasePixels;
1188
1189         if (startPos > lbufWidth)       // Case #4: Begin out, end in, R to L
1190 /*              clippedWidth = startPos - lbufWidth,
1191                 dataClippedWidth = phraseClippedWidth = clippedWidth / phraseWidthToPixels[depth],
1192                 startPos = lbufWidth + (clippedWidth % phraseWidthToPixels[depth]);*/
1193                 clippedWidth = startPos - lbufWidth,
1194                 dataClippedWidth = phraseClippedWidth = clippedWidth / scaledPhrasePixels,
1195                 startPos = lbufWidth + (clippedWidth % scaledPhrasePixels);
1196
1197 extern int op_start_log;
1198 if (op_start_log && clippedWidth != 0)
1199         WriteLog("OP: Clipped line. SP=%i, EP=%i, clip=%u, iwidth=%u, hscale=%02X\n", startPos, endPos, clippedWidth, iwidth, hscale);
1200 if (op_start_log && startPos == 13)
1201 {
1202         WriteLog("OP: Scaled line. SP=%i, EP=%i, clip=%u, iwidth=%u, hscale=%02X, depth=%u, firstPix=%u\n", startPos, endPos, clippedWidth, iwidth, hscale, depth, firstPix);
1203         DumpScaledObject(p0, p1, p2);
1204 }
1205         // If the image is sitting on the line buffer left or right edge, we need to compensate
1206         // by decreasing the image phrase width accordingly.
1207         iwidth -= phraseClippedWidth;
1208
1209         // Also, if we're clipping the phrase we need to make sure we're in the correct part of
1210         // the pixel data.
1211 //      data += phraseClippedWidth * (pitch << 3);
1212         data += dataClippedWidth * (pitch << 3);
1213
1214         // NOTE: When the bitmap is in REFLECT mode, the XPOS marks the *right* side of the
1215         //       bitmap! This makes clipping & etc. MUCH, much easier...!
1216 //      uint32 lbufAddress = 0x1800 + (!in24BPPMode ? leftMargin * 2 : leftMargin * 4);
1217 //      uint32 lbufAddress = 0x1800 + (!in24BPPMode ? startPos * 2 : startPos * 4);
1218         uint32 lbufAddress = 0x1800 + (!in24BPPMode ? startPos * 2 : startPos * 2);
1219         uint8 * currentLineBuffer = &tom_ram_8[lbufAddress];
1220
1221         // Render.
1222
1223 // Hmm. We check above for 24 BPP mode, but don't do anything about it below...
1224 // If we *were* in 24 BPP mode, how would you convert CRY to RGB24? Seems to me
1225 // that if you're in CRY mode then you wouldn't be able to use 24 BPP bitmaps
1226 // anyway.
1227 // This seems to be the case (at least according to the Midsummer docs)...!
1228
1229         if (depth == 0)                                                                 // 1 BPP
1230         {
1231 if (firstPix != 0)
1232         WriteLog("OP: Scaled bitmap @ 1 BPP requesting FIRSTPIX!\n");
1233                 // The LSB of flags is OPFLAG_REFLECT, so sign extend it and or 2 into it.
1234                 int32 lbufDelta = ((int8)((flags << 7) & 0xFF) >> 5) | 0x02;
1235
1236                 int pixCount = 0;
1237                 uint64 pixels = ((uint64)JaguarReadLong(data, OP) << 32) | JaguarReadLong(data + 4, OP);
1238
1239                 while ((int32)iwidth > 0)
1240                 {
1241                         uint8 bits = pixels >> 63;
1242
1243                         if (flagTRANS && bits == 0)
1244                                 ;       // Do nothing...
1245                         else
1246                         {
1247                                 if (!flagRMW)
1248                                         // This is the *only* correct use of endian-dependent code
1249                                         // (i.e., mem-to-mem direct copying)!
1250                                         *(uint16 *)currentLineBuffer = paletteRAM16[index | bits];
1251                                 else
1252                                         *currentLineBuffer =
1253                                                 BLEND_CR(*currentLineBuffer, paletteRAM[(index | bits) << 1]),
1254                                         *(currentLineBuffer + 1) =
1255                                                 BLEND_Y(*(currentLineBuffer + 1), paletteRAM[((index | bits) << 1) + 1]);
1256                         }
1257
1258                         currentLineBuffer += lbufDelta;
1259
1260                         horizontalRemainder -= 0x20;            // Subtract 1.0f in [3.5] fixed point format
1261                         while (horizontalRemainder & 0x80)
1262                         {
1263                                 horizontalRemainder += hscale;
1264                                 pixCount++;
1265                                 pixels <<= 1;
1266                         }
1267
1268                         if (pixCount > 63)
1269                         {
1270                                 int phrasesToSkip = pixCount / 64, pixelShift = pixCount % 64;
1271
1272                                 data += (pitch << 3) * phrasesToSkip;
1273                                 pixels = ((uint64)JaguarReadLong(data, OP) << 32) | JaguarReadLong(data + 4, OP);
1274                                 pixels <<= 1 * pixelShift;
1275                                 iwidth -= phrasesToSkip;
1276                                 pixCount = pixelShift;
1277                         }
1278                 }
1279         }
1280         else if (depth == 1)                                                    // 2 BPP
1281         {
1282 if (firstPix != 0)
1283         WriteLog("OP: Scaled bitmap @ 2 BPP requesting FIRSTPIX!\n");
1284                 index &= 0xFC;                                                          // Top six bits form CLUT index
1285                 // The LSB is OPFLAG_REFLECT, so sign extend it and or 2 into it.
1286                 int32 lbufDelta = ((int8)((flags << 7) & 0xFF) >> 5) | 0x02;
1287
1288                 int pixCount = 0;
1289                 uint64 pixels = ((uint64)JaguarReadLong(data, OP) << 32) | JaguarReadLong(data + 4, OP);
1290
1291                 while ((int32)iwidth > 0)
1292                 {
1293                         uint8 bits = pixels >> 62;
1294
1295                         if (flagTRANS && bits == 0)
1296                                 ;       // Do nothing...
1297                         else
1298                         {
1299                                 if (!flagRMW)
1300                                         // This is the *only* correct use of endian-dependent code
1301                                         // (i.e., mem-to-mem direct copying)!
1302                                         *(uint16 *)currentLineBuffer = paletteRAM16[index | bits];
1303                                 else
1304                                         *currentLineBuffer =
1305                                                 BLEND_CR(*currentLineBuffer, paletteRAM[(index | bits) << 1]),
1306                                         *(currentLineBuffer + 1) =
1307                                                 BLEND_Y(*(currentLineBuffer + 1), paletteRAM[((index | bits) << 1) + 1]);
1308                         }
1309
1310                         currentLineBuffer += lbufDelta;
1311
1312                         horizontalRemainder -= 0x20;            // Subtract 1.0f in [3.5] fixed point format
1313                         while (horizontalRemainder & 0x80)
1314                         {
1315                                 horizontalRemainder += hscale;
1316                                 pixCount++;
1317                                 pixels <<= 2;
1318                         }
1319
1320                         if (pixCount > 31)
1321                         {
1322                                 int phrasesToSkip = pixCount / 32, pixelShift = pixCount % 32;
1323
1324                                 data += (pitch << 3) * phrasesToSkip;
1325                                 pixels = ((uint64)JaguarReadLong(data, OP) << 32) | JaguarReadLong(data + 4, OP);
1326                                 pixels <<= 2 * pixelShift;
1327                                 iwidth -= phrasesToSkip;
1328                                 pixCount = pixelShift;
1329                         }
1330                 }
1331         }
1332         else if (depth == 2)                                                    // 4 BPP
1333         {
1334 if (firstPix != 0)
1335         WriteLog("OP: Scaled bitmap @ 4 BPP requesting FIRSTPIX!\n");
1336                 index &= 0xF0;                                                          // Top four bits form CLUT index
1337                 // The LSB is OPFLAG_REFLECT, so sign extend it and or 2 into it.
1338                 int32 lbufDelta = ((int8)((flags << 7) & 0xFF) >> 5) | 0x02;
1339
1340                 int pixCount = 0;
1341                 uint64 pixels = ((uint64)JaguarReadLong(data, OP) << 32) | JaguarReadLong(data + 4, OP);
1342
1343                 while ((int32)iwidth > 0)
1344                 {
1345                         uint8 bits = pixels >> 60;
1346
1347                         if (flagTRANS && bits == 0)
1348                                 ;       // Do nothing...
1349                         else
1350                         {
1351                                 if (!flagRMW)
1352                                         // This is the *only* correct use of endian-dependent code
1353                                         // (i.e., mem-to-mem direct copying)!
1354                                         *(uint16 *)currentLineBuffer = paletteRAM16[index | bits];
1355                                 else
1356                                         *currentLineBuffer =
1357                                                 BLEND_CR(*currentLineBuffer, paletteRAM[(index | bits) << 1]),
1358                                         *(currentLineBuffer + 1) =
1359                                                 BLEND_Y(*(currentLineBuffer + 1), paletteRAM[((index | bits) << 1) + 1]);
1360                         }
1361
1362                         currentLineBuffer += lbufDelta;
1363
1364                         horizontalRemainder -= 0x20;            // Subtract 1.0f in [3.5] fixed point format
1365                         while (horizontalRemainder & 0x80)
1366                         {
1367                                 horizontalRemainder += hscale;
1368                                 pixCount++;
1369                                 pixels <<= 4;
1370                         }
1371
1372                         if (pixCount > 15)
1373                         {
1374                                 int phrasesToSkip = pixCount / 16, pixelShift = pixCount % 16;
1375
1376                                 data += (pitch << 3) * phrasesToSkip;
1377                                 pixels = ((uint64)JaguarReadLong(data, OP) << 32) | JaguarReadLong(data + 4, OP);
1378                                 pixels <<= 4 * pixelShift;
1379                                 iwidth -= phrasesToSkip;
1380                                 pixCount = pixelShift;
1381                         }
1382                 }
1383         }
1384         else if (depth == 3)                                                    // 8 BPP
1385         {
1386 if (firstPix)
1387         WriteLog("OP: Scaled bitmap @ 8 BPP requesting FIRSTPIX! (fp=%u)\n", firstPix);
1388                 // The LSB is OPFLAG_REFLECT, so sign extend it and or 2 into it.
1389                 int32 lbufDelta = ((int8)((flags << 7) & 0xFF) >> 5) | 0x02;
1390
1391                 int pixCount = 0;
1392                 uint64 pixels = ((uint64)JaguarReadLong(data, OP) << 32) | JaguarReadLong(data + 4, OP);
1393
1394                 while ((int32)iwidth > 0)
1395                 {
1396                         uint8 bits = pixels >> 56;
1397
1398                         if (flagTRANS && bits == 0)
1399                                 ;       // Do nothing...
1400                         else
1401                         {
1402                                 if (!flagRMW)
1403                                         // This is the *only* correct use of endian-dependent code
1404                                         // (i.e., mem-to-mem direct copying)!
1405                                         *(uint16 *)currentLineBuffer = paletteRAM16[bits];
1406                                 else
1407                                         *currentLineBuffer =
1408                                                 BLEND_CR(*currentLineBuffer, paletteRAM[bits << 1]),
1409                                         *(currentLineBuffer + 1) =
1410                                                 BLEND_Y(*(currentLineBuffer + 1), paletteRAM[(bits << 1) + 1]);
1411                         }
1412
1413                         currentLineBuffer += lbufDelta;
1414
1415                         horizontalRemainder -= 0x20;            // Subtract 1.0f in [3.5] fixed point format
1416                         while (horizontalRemainder & 0x80)
1417                         {
1418                                 horizontalRemainder += hscale;
1419                                 pixCount++;
1420                                 pixels <<= 8;
1421                         }
1422
1423                         if (pixCount > 7)
1424                         {
1425                                 int phrasesToSkip = pixCount / 8, pixelShift = pixCount % 8;
1426
1427                                 data += (pitch << 3) * phrasesToSkip;
1428                                 pixels = ((uint64)JaguarReadLong(data, OP) << 32) | JaguarReadLong(data + 4, OP);
1429                                 pixels <<= 8 * pixelShift;
1430                                 iwidth -= phrasesToSkip;
1431                                 pixCount = pixelShift;
1432                         }
1433                 }
1434         }
1435         else if (depth == 4)                                                    // 16 BPP
1436         {
1437 if (firstPix != 0)
1438         WriteLog("OP: Scaled bitmap @ 16 BPP requesting FIRSTPIX!\n");
1439                 // The LSB is OPFLAG_REFLECT, so sign extend it and OR 2 into it.
1440                 int32 lbufDelta = ((int8)((flags << 7) & 0xFF) >> 5) | 0x02;
1441
1442                 int pixCount = 0;
1443                 uint64 pixels = ((uint64)JaguarReadLong(data, OP) << 32) | JaguarReadLong(data + 4, OP);
1444
1445                 while ((int32)iwidth > 0)
1446                 {
1447                         uint8 bitsHi = pixels >> 56, bitsLo = pixels >> 48;
1448
1449                         if (flagTRANS && (bitsLo | bitsHi) == 0)
1450                                 ;       // Do nothing...
1451                         else
1452                         {
1453                                 if (!flagRMW)
1454                                         *currentLineBuffer = bitsHi,
1455                                         *(currentLineBuffer + 1) = bitsLo;
1456                                 else
1457                                         *currentLineBuffer =
1458                                                 BLEND_CR(*currentLineBuffer, bitsHi),
1459                                         *(currentLineBuffer + 1) =
1460                                                 BLEND_Y(*(currentLineBuffer + 1), bitsLo);
1461                         }
1462
1463                         currentLineBuffer += lbufDelta;
1464
1465                         horizontalRemainder -= 0x20;            // Subtract 1.0f in [3.5] fixed point format
1466                         while (horizontalRemainder & 0x80)
1467                         {
1468                                 horizontalRemainder += hscale;
1469                                 pixCount++;
1470                                 pixels <<= 16;
1471                         }
1472
1473                         if (pixCount > 3)
1474                         {
1475                                 int phrasesToSkip = pixCount / 4, pixelShift = pixCount % 4;
1476
1477                                 data += (pitch << 3) * phrasesToSkip;
1478                                 pixels = ((uint64)JaguarReadLong(data, OP) << 32) | JaguarReadLong(data + 4, OP);
1479                                 pixels <<= 16 * pixelShift;
1480
1481                                 iwidth -= phrasesToSkip;
1482
1483                                 pixCount = pixelShift;
1484                         }
1485                 }
1486         }
1487         else if (depth == 5)                                                    // 24 BPP
1488         {
1489 //I'm not sure that you can scale a 24 BPP bitmap properly--the JTRM seem to indicate as much.
1490 WriteLog("OP: Writing 24 BPP scaled bitmap!\n");
1491 if (firstPix != 0)
1492         WriteLog("OP: Scaled bitmap @ 24 BPP requesting FIRSTPIX!\n");
1493                 // Not sure, but I think RMW only works with 16 BPP and below, and only in CRY mode...
1494                 // The LSB is OPFLAG_REFLECT, so sign extend it and or 4 into it.
1495                 int32 lbufDelta = ((int8)((flags << 7) & 0xFF) >> 4) | 0x04;
1496
1497                 while (iwidth--)
1498                 {
1499                         // Fetch phrase...
1500                         uint64 pixels = ((uint64)JaguarReadLong(data, OP) << 32) | JaguarReadLong(data + 4, OP);
1501                         data += pitch << 3;                                             // Multiply pitch * 8 (optimize: precompute this value)
1502
1503                         for(int i=0; i<2; i++)
1504                         {
1505                                 uint8 bits3 = pixels >> 56, bits2 = pixels >> 48,
1506                                         bits1 = pixels >> 40, bits0 = pixels >> 32;
1507
1508                                 if (flagTRANS && (bits3 | bits2 | bits1 | bits0) == 0)
1509                                         ;       // Do nothing...
1510                                 else
1511                                         *currentLineBuffer = bits3,
1512                                         *(currentLineBuffer + 1) = bits2,
1513                                         *(currentLineBuffer + 2) = bits1,
1514                                         *(currentLineBuffer + 3) = bits0;
1515
1516                                 currentLineBuffer += lbufDelta;
1517                                 pixels <<= 32;
1518                         }
1519                 }
1520         }
1521 /*if (depth == 3 && startPos == 13)
1522 {
1523 if (op_start_log)
1524 WriteLog("OP: Writing in the margins...\n");
1525         for(int i=0; i<100*2; i+=2)
1526 //      for(int i=0; i<14*2; i+=2)
1527                 tom_ram_8[0x1800 + i] = 0xFF,
1528                 tom_ram_8[0x1800 + i + 1] = 0xFF;
1529 }*/
1530 //      uint32 lbufAddress = 0x1800 + (!in24BPPMode ? startPos * 2 : startPos * 4);
1531 //      uint8 * currentLineBuffer = &tom_ram_8[lbufAddress];
1532 }