| 1 | /*\r |
| 2 | * assembly optimized versions of most funtions from draw2.c\r |
| 3 | * (C) notaz, 2006-2008\r |
| 4 | *\r |
| 5 | * This work is licensed under the terms of MAME license.\r |
| 6 | * See COPYING file in the top-level directory.\r |
| 7 | *\r |
| 8 | * this is highly specialized, be careful if changing related C code!\r |
| 9 | */\r |
| 10 | \r |
| 11 | #include "pico_int_o32.h"\r |
| 12 | \r |
| 13 | @ define these constants in your include file:\r |
| 14 | @ .equiv START_ROW, 1\r |
| 15 | @ .equiv END_ROW, 27\r |
| 16 | @ one row means 8 pixels. If above example was used, (27-1)*8=208 lines would be rendered.\r |
| 17 | #ifndef START_ROW\r |
| 18 | #define START_ROW 0\r |
| 19 | #endif\r |
| 20 | #ifndef END_ROW\r |
| 21 | #define END_ROW 28\r |
| 22 | #endif\r |
| 23 | \r |
| 24 | .text\r |
| 25 | .align 2\r |
| 26 | \r |
| 27 | @ void BackFillFull(void *dst, int reg7)\r |
| 28 | \r |
| 29 | .global BackFillFull\r |
| 30 | \r |
| 31 | BackFillFull:\r |
| 32 | stmfd sp!, {r4-r9,lr}\r |
| 33 | \r |
| 34 | add lr, r0, #328*8\r |
| 35 | mov r0, r1, lsl #26\r |
| 36 | mov r0, r0, lsr #26\r |
| 37 | \r |
| 38 | orr r0, r0, r0, lsl #8\r |
| 39 | orr r0, r0, r0, lsl #16\r |
| 40 | \r |
| 41 | mov r1, r0 @ 25 opcodes wasted?\r |
| 42 | mov r2, r0\r |
| 43 | mov r3, r0\r |
| 44 | mov r4, r0\r |
| 45 | mov r5, r0\r |
| 46 | mov r6, r0\r |
| 47 | mov r7, r0\r |
| 48 | mov r8, r0\r |
| 49 | mov r9, r0\r |
| 50 | \r |
| 51 | mov r12, #(END_ROW-START_ROW)*8\r |
| 52 | \r |
| 53 | @ go go go!\r |
| 54 | .bff_loop:\r |
| 55 | add lr, lr, #8\r |
| 56 | subs r12, r12, #1\r |
| 57 | \r |
| 58 | stmia lr!, {r0-r9} @ 10*4*8\r |
| 59 | stmia lr!, {r0-r9}\r |
| 60 | stmia lr!, {r0-r9}\r |
| 61 | stmia lr!, {r0-r9}\r |
| 62 | stmia lr!, {r0-r9}\r |
| 63 | stmia lr!, {r0-r9}\r |
| 64 | stmia lr!, {r0-r9}\r |
| 65 | stmia lr!, {r0-r9}\r |
| 66 | \r |
| 67 | bne .bff_loop\r |
| 68 | \r |
| 69 | ldmfd sp!, {r4-r9,lr}\r |
| 70 | bx lr\r |
| 71 | \r |
| 72 | .pool\r |
| 73 | \r |
| 74 | @ -------- some macros --------\r |
| 75 | \r |
| 76 | @ helpers\r |
| 77 | .macro add_c24 d s c\r |
| 78 | add \d, \s, #(\c & 0x00ff00)\r |
| 79 | .if \c & 0x0000ff\r |
| 80 | add \d, \d, #(\c & 0x0000ff)\r |
| 81 | .endif\r |
| 82 | .if \c & 0xff0000\r |
| 83 | add \d, \d, #(\c & 0xff0000)\r |
| 84 | .endif\r |
| 85 | .endm\r |
| 86 | \r |
| 87 | @ TileLineSinglecol (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r0: pixels8_old\r |
| 88 | .macro TileLineSinglecol notsinglecol=0\r |
| 89 | and r2, r2, #0xf @ #0x0000000f\r |
| 90 | .if !\notsinglecol\r |
| 91 | cmp r2, r0, lsr #28 @ if these don't match,\r |
| 92 | bicne r9, r9, #2 @ it is a sign that whole tile is not singlecolor (only it's lines may be)\r |
| 93 | .endif\r |
| 94 | orr r4, r3, r2\r |
| 95 | orr r4, r4, r4, lsl #8\r |
| 96 | \r |
| 97 | tst r1, #1 @ not aligned?\r |
| 98 | strneb r4, [r1], #1\r |
| 99 | streqh r4, [r1], #2\r |
| 100 | strh r4, [r1], #2\r |
| 101 | strh r4, [r1], #2\r |
| 102 | strh r4, [r1], #2\r |
| 103 | strneb r4, [r1], #1 @ have a remaining unaligned pixel?\r |
| 104 | sub r1, r1, #8\r |
| 105 | .if !\notsinglecol\r |
| 106 | mov r0, #0xf\r |
| 107 | orr r0, r0, r2, lsl #28 @ we will need the old palindex later\r |
| 108 | .endif\r |
| 109 | .endm\r |
| 110 | \r |
| 111 | @ TileNorm (r1=pdest, r2=pixels8, r3=pal) r0,r4: scratch\r |
| 112 | .macro TileLineNorm\r |
| 113 | ands r4, r0, r2, lsr #12 @ #0x0000f000\r |
| 114 | orrne r4, r3, r4\r |
| 115 | strneb r4, [r1]\r |
| 116 | ands r4, r0, r2, lsr #8 @ #0x00000f00\r |
| 117 | orrne r4, r3, r4\r |
| 118 | strneb r4, [r1,#1]\r |
| 119 | ands r4, r0, r2, lsr #4 @ #0x000000f0\r |
| 120 | orrne r4, r3, r4\r |
| 121 | strneb r4, [r1,#2]\r |
| 122 | ands r4, r0, r2 @ #0x0000000f\r |
| 123 | orrne r4, r3, r4\r |
| 124 | strneb r4, [r1,#3]\r |
| 125 | ands r4, r0, r2, lsr #28 @ #0xf0000000\r |
| 126 | orrne r4, r3, r4\r |
| 127 | strneb r4, [r1,#4]\r |
| 128 | ands r4, r0, r2, lsr #24 @ #0x0f000000\r |
| 129 | orrne r4, r3, r4\r |
| 130 | strneb r4, [r1,#5]\r |
| 131 | ands r4, r0, r2, lsr #20 @ #0x00f00000\r |
| 132 | orrne r4, r3, r4\r |
| 133 | strneb r4, [r1,#6]\r |
| 134 | ands r4, r0, r2, lsr #16 @ #0x000f0000\r |
| 135 | orrne r4, r3, r4\r |
| 136 | strneb r4, [r1,#7]\r |
| 137 | .endm\r |
| 138 | \r |
| 139 | @ TileFlip (r1=pdest, r2=pixels8, r3=pal) r0,r4: scratch\r |
| 140 | .macro TileLineFlip\r |
| 141 | ands r4, r0, r2, lsr #16 @ #0x000f0000\r |
| 142 | orrne r4, r3, r4\r |
| 143 | strneb r4, [r1]\r |
| 144 | ands r4, r0, r2, lsr #20 @ #0x00f00000\r |
| 145 | orrne r4, r3, r4\r |
| 146 | strneb r4, [r1,#1]\r |
| 147 | ands r4, r0, r2, lsr #24 @ #0x0f000000\r |
| 148 | orrne r4, r3, r4\r |
| 149 | strneb r4, [r1,#2]\r |
| 150 | ands r4, r0, r2, lsr #28 @ #0xf0000000\r |
| 151 | orrne r4, r3, r4\r |
| 152 | strneb r4, [r1,#3]\r |
| 153 | ands r4, r0, r2 @ #0x0000000f\r |
| 154 | orrne r4, r3, r4\r |
| 155 | strneb r4, [r1,#4]\r |
| 156 | ands r4, r0, r2, lsr #4 @ #0x000000f0\r |
| 157 | orrne r4, r3, r4\r |
| 158 | strneb r4, [r1,#5]\r |
| 159 | ands r4, r0, r2, lsr #8 @ #0x00000f00\r |
| 160 | orrne r4, r3, r4\r |
| 161 | strneb r4, [r1,#6]\r |
| 162 | ands r4, r0, r2, lsr #12 @ #0x0000f000\r |
| 163 | orrne r4, r3, r4\r |
| 164 | strneb r4, [r1,#7]\r |
| 165 | .endm\r |
| 166 | \r |
| 167 | @ Tile (r1=pdest, r3=pal, r9=prevcode, r10=Pico.vram) r2,r4,r7: scratch, r0=0xf\r |
| 168 | .macro Tile hflip vflip\r |
| 169 | mov r7, r9, lsl #13 @ r9=code<<8; addr=(code&0x7ff)<<4;\r |
| 170 | add r7, r10, r7, lsr #16\r |
| 171 | orr r9, r9, #3 @ emptytile=singlecolor=1, r9 must be <code_16> 00000xxx\r |
| 172 | .if \vflip\r |
| 173 | @ we read tilecodes in reverse order if we have vflip\r |
| 174 | add r7, r7, #8*4\r |
| 175 | .endif\r |
| 176 | @ loop through 8 lines\r |
| 177 | orr r9, r9, #(7<<24)\r |
| 178 | b 1f @ loop_enter\r |
| 179 | \r |
| 180 | 0: @ singlecol_loop\r |
| 181 | subs r9, r9, #(1<<24)\r |
| 182 | add r1, r1, #328 @ set pointer to next line\r |
| 183 | bmi 8f @ loop_exit with r0 restore\r |
| 184 | 1:\r |
| 185 | .if \vflip\r |
| 186 | ldr r2, [r7, #-4]! @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels\r |
| 187 | .else\r |
| 188 | ldr r2, [r7], #4\r |
| 189 | .endif\r |
| 190 | tst r2, r2\r |
| 191 | beq 2f @ empty line\r |
| 192 | bic r9, r9, #1\r |
| 193 | cmp r2, r2, ror #4\r |
| 194 | bne 3f @ not singlecolor\r |
| 195 | TileLineSinglecol\r |
| 196 | b 0b\r |
| 197 | \r |
| 198 | 2:\r |
| 199 | bic r9, r9, #2\r |
| 200 | 2: @ empty_loop\r |
| 201 | subs r9, r9, #(1<<24)\r |
| 202 | add r1, r1, #328 @ set pointer to next line\r |
| 203 | bmi 8f @ loop_exit with r0 restore\r |
| 204 | .if \vflip\r |
| 205 | ldr r2, [r7, #-4]! @ next pack\r |
| 206 | .else\r |
| 207 | ldr r2, [r7], #4\r |
| 208 | .endif\r |
| 209 | mov r0, #0xf @ singlecol_loop might have messed r0\r |
| 210 | tst r2, r2\r |
| 211 | beq 2b\r |
| 212 | \r |
| 213 | bic r9, r9, #3 @ if we are here, it means we have empty and not empty line\r |
| 214 | b 5f\r |
| 215 | \r |
| 216 | 3: @ not empty, not singlecol\r |
| 217 | mov r0, #0xf\r |
| 218 | bic r9, r9, #3\r |
| 219 | b 6f\r |
| 220 | \r |
| 221 | 4: @ not empty, not singlecol loop\r |
| 222 | subs r9, r9, #(1<<24)\r |
| 223 | add r1, r1, #328 @ set pointer to next line\r |
| 224 | bmi 9f @ loop_exit\r |
| 225 | .if \vflip\r |
| 226 | ldr r2, [r7, #-4]! @ next pack\r |
| 227 | .else\r |
| 228 | ldr r2, [r7], #4\r |
| 229 | .endif\r |
| 230 | tst r2, r2\r |
| 231 | beq 4b @ empty line\r |
| 232 | 5:\r |
| 233 | cmp r2, r2, ror #4\r |
| 234 | beq 7f @ singlecolor line\r |
| 235 | 6:\r |
| 236 | .if \hflip\r |
| 237 | TileLineFlip\r |
| 238 | .else\r |
| 239 | TileLineNorm\r |
| 240 | .endif\r |
| 241 | b 4b\r |
| 242 | 7:\r |
| 243 | TileLineSinglecol 1\r |
| 244 | b 4b\r |
| 245 | \r |
| 246 | 8:\r |
| 247 | mov r0, #0xf\r |
| 248 | 9: @ loop_exit\r |
| 249 | add r9, r9, #(1<<24) @ fix r9\r |
| 250 | sub r1, r1, #328*8 @ restore pdest pointer\r |
| 251 | .endm\r |
| 252 | \r |
| 253 | \r |
| 254 | @ TileLineSinglecolAl (r1=pdest, r4,r7=color)\r |
| 255 | .macro TileLineSinglecolAl0\r |
| 256 | stmia r1!, {r4,r7}\r |
| 257 | add r1, r1, #320\r |
| 258 | .endm\r |
| 259 | \r |
| 260 | .macro TileLineSinglecolAl1\r |
| 261 | strb r4, [r1], #1\r |
| 262 | strh r4, [r1], #2\r |
| 263 | str r4, [r1], #4\r |
| 264 | strb r4, [r1], #1+320\r |
| 265 | @ add r1, r1, #320\r |
| 266 | .endm\r |
| 267 | \r |
| 268 | .macro TileLineSinglecolAl2\r |
| 269 | strh r4, [r1], #2\r |
| 270 | str r4, [r1], #4\r |
| 271 | strh r4, [r1], #2\r |
| 272 | add r1, r1, #320\r |
| 273 | .endm\r |
| 274 | \r |
| 275 | .macro TileLineSinglecolAl3\r |
| 276 | strb r4, [r1], #1\r |
| 277 | str r4, [r1], #4\r |
| 278 | strh r4, [r1], #2\r |
| 279 | strb r4, [r1], #1+320\r |
| 280 | @ add r1, r1, #320\r |
| 281 | .endm\r |
| 282 | \r |
| 283 | @ TileSinglecol (r1=pdest, r2=pixels8, r3=pal) r4,r7: scratch, r0=0xf\r |
| 284 | @ kaligned==1, if dest is always aligned\r |
| 285 | .macro TileSinglecol kaligned=0\r |
| 286 | and r4, r2, #0xf @ we assume we have good r2 from previous time\r |
| 287 | orr r4, r4, r3\r |
| 288 | orr r4, r4, r4, lsl #8\r |
| 289 | orr r4, r4, r4, lsl #16\r |
| 290 | mov r7, r4\r |
| 291 | \r |
| 292 | .if !\kaligned\r |
| 293 | tst r1, #2 @ not aligned?\r |
| 294 | bne 2f\r |
| 295 | tst r1, #1\r |
| 296 | bne 1f\r |
| 297 | .endif\r |
| 298 | \r |
| 299 | TileLineSinglecolAl0\r |
| 300 | TileLineSinglecolAl0\r |
| 301 | TileLineSinglecolAl0\r |
| 302 | TileLineSinglecolAl0\r |
| 303 | TileLineSinglecolAl0\r |
| 304 | TileLineSinglecolAl0\r |
| 305 | TileLineSinglecolAl0\r |
| 306 | TileLineSinglecolAl0\r |
| 307 | \r |
| 308 | .if !\kaligned\r |
| 309 | b 4f\r |
| 310 | 1:\r |
| 311 | TileLineSinglecolAl1\r |
| 312 | TileLineSinglecolAl1\r |
| 313 | TileLineSinglecolAl1\r |
| 314 | TileLineSinglecolAl1\r |
| 315 | TileLineSinglecolAl1\r |
| 316 | TileLineSinglecolAl1\r |
| 317 | TileLineSinglecolAl1\r |
| 318 | TileLineSinglecolAl1\r |
| 319 | b 4f\r |
| 320 | \r |
| 321 | 2:\r |
| 322 | tst r1, #1\r |
| 323 | bne 3f\r |
| 324 | \r |
| 325 | TileLineSinglecolAl2\r |
| 326 | TileLineSinglecolAl2\r |
| 327 | TileLineSinglecolAl2\r |
| 328 | TileLineSinglecolAl2\r |
| 329 | TileLineSinglecolAl2\r |
| 330 | TileLineSinglecolAl2\r |
| 331 | TileLineSinglecolAl2\r |
| 332 | TileLineSinglecolAl2\r |
| 333 | b 4f\r |
| 334 | \r |
| 335 | 3:\r |
| 336 | TileLineSinglecolAl3\r |
| 337 | TileLineSinglecolAl3\r |
| 338 | TileLineSinglecolAl3\r |
| 339 | TileLineSinglecolAl3\r |
| 340 | TileLineSinglecolAl3\r |
| 341 | TileLineSinglecolAl3\r |
| 342 | TileLineSinglecolAl3\r |
| 343 | TileLineSinglecolAl3\r |
| 344 | \r |
| 345 | 4:\r |
| 346 | .endif\r |
| 347 | sub r1, r1, #328*8 @ restore pdest pointer\r |
| 348 | .endm\r |
| 349 | \r |
| 350 | \r |
| 351 | \r |
| 352 | @ DrawLayerTiles(*hcache, *scrpos, (cells<<24)|(nametab<<9)|(vscroll&0x3ff)<<11|(shift[width]<<8)|planeend, (ymask<<24)|(planestart<<16)|[htab||hscroll]\r |
| 353 | \r |
| 354 | @ void DrawLayerFull(int plane, int *hcache, int planestart, int planeend,\r |
| 355 | @ struct PicoEState *est)\r |
| 356 | \r |
| 357 | .global DrawLayerFull\r |
| 358 | \r |
| 359 | DrawLayerFull:\r |
| 360 | ldr r12,[sp] @ est\r |
| 361 | stmfd sp!, {r4-r11,lr}\r |
| 362 | \r |
| 363 | mov r6, r1 @ hcache\r |
| 364 | \r |
| 365 | ldr r11, [r12, #OFS_EST_Pico]\r |
| 366 | ldr r10, [r12, #OFS_EST_PicoMem_vram]\r |
| 367 | ldrb r5, [r11, #OFS_Pico_video_reg+13] @ pvid->reg[13]\r |
| 368 | ldrb r7, [r11, #OFS_Pico_video_reg+11]\r |
| 369 | \r |
| 370 | sub lr, r3, r2\r |
| 371 | and lr, lr, #0x00ff0000 @ lr=cells\r |
| 372 | \r |
| 373 | mov r5, r5, lsl #10 @ htab=pvid->reg[13]<<9; (halfwords)\r |
| 374 | add r5, r5, r0, lsl #1 @ htab+=plane\r |
| 375 | bic r5, r5, #0x00ff0000 @ just in case\r |
| 376 | \r |
| 377 | tst r7, #3 @ full screen scroll? (if ==0)\r |
| 378 | ldrb r7, [r11, #OFS_Pico_video_reg+16] @ ??hh??ww\r |
| 379 | ldreqh r5, [r10, r5]\r |
| 380 | biceq r5, r5, #0x0000fc00 @ r5=hscroll (0-0x3ff)\r |
| 381 | movne r5, r5, lsr #1\r |
| 382 | orrne r5, r5, #0x8000 @ this marks that we have htab pointer, not hscroll here\r |
| 383 | \r |
| 384 | and r8, r7, #3\r |
| 385 | \r |
| 386 | orr r5, r5, r7, lsl #1+24\r |
| 387 | orr r5, r5, #0x1f000000\r |
| 388 | cmp r8, #1\r |
| 389 | biclt r5, r5, #0x80000000\r |
| 390 | biceq r5, r5, #0xc0000000\r |
| 391 | bicgt r5, r5, #0xe0000000\r |
| 392 | \r |
| 393 | mov r9, r2, lsl #24\r |
| 394 | orr r5, r5, r9, lsr #8 @ r5=(ymask<<24)|(trow<<16)|[htab||hscroll]\r |
| 395 | \r |
| 396 | add r4, r8, #5\r |
| 397 | cmp r4, #7\r |
| 398 | subge r4, r4, #1 @ r4=shift[width] (5,6,6,7)\r |
| 399 | \r |
| 400 | orr lr, lr, r4 \r |
| 401 | orr lr, lr, r3, lsl #24 @ lr=(planeend<<24)|(cells<<16)|shift[width]\r |
| 402 | \r |
| 403 | @ calculate xmask:\r |
| 404 | mov r8, r8, lsl #24+5\r |
| 405 | orr r8, r8, #0x1f000000\r |
| 406 | \r |
| 407 | @ Find name table:\r |
| 408 | tst r0, r0\r |
| 409 | ldreqb r4, [r11, #OFS_Pico_video_reg+2]\r |
| 410 | moveq r4, r4, lsr #3\r |
| 411 | ldrneb r4, [r11, #OFS_Pico_video_reg+4]\r |
| 412 | and r4, r4, #7\r |
| 413 | orr lr, lr, r4, lsl #13 @ lr|=nametab_bits{3}<<13\r |
| 414 | \r |
| 415 | ldr r11,[sp, #9*4] @ est\r |
| 416 | sub r4, r9, #(START_ROW<<24)\r |
| 417 | ldr r11, [r11, #OFS_EST_Draw2FB]\r |
| 418 | mov r4, r4, asr #24\r |
| 419 | mov r7, #328*8\r |
| 420 | mla r11, r4, r7, r11 @ scrpos+=8*328*(planestart-START_ROW);\r |
| 421 | \r |
| 422 | @ Get vertical scroll value:\r |
| 423 | add_c24 r7, r10, (OFS_PMEM_vsram-OFS_PMEM_vram)\r |
| 424 | ldr r7, [r7]\r |
| 425 | tst r0, r0\r |
| 426 | moveq r7, r7, lsl #22\r |
| 427 | movne r7, r7, lsl #6\r |
| 428 | mov r7, r7, lsr #22 @ r7=vscroll (10 bits)\r |
| 429 | \r |
| 430 | orr lr, lr, r7, lsl #3\r |
| 431 | mov lr, lr, ror #24 @ packed: cccccccc nnnvvvvv vvvvvsss pppppppp: cells, nametab, vscroll, shift[width], planeend\r |
| 432 | \r |
| 433 | ands r7, r7, #7\r |
| 434 | addne lr, lr, #1 @ we have vertically clipped tiles due to vscroll, so we need 1 more row\r |
| 435 | \r |
| 436 | rsb r7, r7, #8\r |
| 437 | str r7, [r6], #4 @ push y-offset to tilecache\r |
| 438 | mov r4, #328\r |
| 439 | mla r11, r4, r7, r11 @ scrpos+=(8-(vscroll&7))*328;\r |
| 440 | \r |
| 441 | mov r9, #0xff000000 @ r9=(prevcode<<8)|flags: 1~tile empty, 2~tile singlecolor\r |
| 442 | \r |
| 443 | .rtrloop_outer:\r |
| 444 | mov r4, lr, lsl #11\r |
| 445 | mov r4, r4, lsr #25 @ r4=vscroll>>3 (7 bits)\r |
| 446 | add r4, r4, r5, lsr #16 @ +trow\r |
| 447 | and r4, r4, r5, lsr #24 @ &=ymask\r |
| 448 | mov r7, lr, lsr #8\r |
| 449 | and r7, r7, #7 @ shift[width]\r |
| 450 | mov r0, lr, lsr #9\r |
| 451 | and r0, r0, #0x7000 @ nametab\r |
| 452 | add r12,r0, r4, lsl r7 @ nametab_row = nametab + (((trow+(vscroll>>3))&ymask)<<shift[width]); \r |
| 453 | \r |
| 454 | mov r4, lr, lsr #24\r |
| 455 | orr r12,r12,r4, lsl #23\r |
| 456 | mov r12,r12,lsl #1 @ (nametab_row|(cells<<24)) (halfword compliant)\r |
| 457 | \r |
| 458 | @ htab?\r |
| 459 | tst r5, #0x8000\r |
| 460 | moveq r7, r5, lsl #22 @ hscroll (0-3FFh)\r |
| 461 | moveq r7, r7, lsr #22\r |
| 462 | beq .rtr_hscroll_done\r |
| 463 | \r |
| 464 | @ get hscroll from htab\r |
| 465 | mov r7, r5, lsl #17\r |
| 466 | ands r4, r5, #0x00ff0000\r |
| 467 | add r7, r7, r4, lsl #5 @ +=trow<<4\r |
| 468 | andne r4, lr, #0x3800\r |
| 469 | subne r7, r7, r4, lsl #7 @ if(trow) htaddr-=(vscroll&7)<<1;\r |
| 470 | mov r7, r7, lsr #16 @ halfwords\r |
| 471 | ldrh r7, [r10, r7]\r |
| 472 | \r |
| 473 | .rtr_hscroll_done:\r |
| 474 | and r8, r8, #0xff000000\r |
| 475 | rsb r4, r7, #0 @ r4=tilex=(-ts->hscroll)>>3\r |
| 476 | mov r4, r4, asr #3\r |
| 477 | and r4, r4, #0xff\r |
| 478 | orr r8, r8, r4 @ r8=(xmask<<24)|tilex\r |
| 479 | \r |
| 480 | sub r7, r7, #1\r |
| 481 | and r7, r7, #7\r |
| 482 | add r7, r7, #1 @ r7=dx=((ts->hscroll-1)&7)+1\r |
| 483 | \r |
| 484 | cmp r7, #8\r |
| 485 | subeq r12,r12, #0x01000000 @ we will loop cells+1 times, so loop less when there is no hscroll\r |
| 486 | \r |
| 487 | add r1, r11, r7 @ r1=pdest\r |
| 488 | mov r0, #0xf\r |
| 489 | b .rtrloop_enter\r |
| 490 | \r |
| 491 | @ r4 & r7 are scratch in this loop\r |
| 492 | .rtrloop: @ 40-41 times\r |
| 493 | add r1, r1, #8\r |
| 494 | subs r12,r12, #0x01000000\r |
| 495 | add r8, r8, #1\r |
| 496 | bmi .rtrloop_exit\r |
| 497 | \r |
| 498 | .rtrloop_enter:\r |
| 499 | and r7, r8, r8, lsr #24\r |
| 500 | add r7, r10, r7, lsl #1\r |
| 501 | bic r4, r12, #0xff000000 @ Pico.vram[nametab_row+(tilex&xmask)];\r |
| 502 | ldrh r7, [r7, r4] @ r7=code (int, but from unsigned, no sign extend)\r |
| 503 | \r |
| 504 | tst r7, #0x8000\r |
| 505 | bne .rtr_hiprio\r |
| 506 | \r |
| 507 | cmp r7, r9, lsr #8\r |
| 508 | bne .rtr_notsamecode\r |
| 509 | @ we know stuff about this tile already\r |
| 510 | tst r9, #1\r |
| 511 | bne .rtrloop @ empty tile\r |
| 512 | tst r9, #2\r |
| 513 | bne .rtr_singlecolor @ singlecolor tile\r |
| 514 | b .rtr_samecode\r |
| 515 | \r |
| 516 | .rtr_notsamecode:\r |
| 517 | and r4, r9, #0x600000\r |
| 518 | mov r9, r7, lsl #8 @ remember new code\r |
| 519 | \r |
| 520 | @ update cram\r |
| 521 | and r7, r7, #0x6000\r |
| 522 | mov r3, r7, asr #9 @ r3=pal=((code&0x6000)>>9);\r |
| 523 | \r |
| 524 | .rtr_samecode:\r |
| 525 | tst r9, #0x100000 @ vflip?\r |
| 526 | bne .rtr_vflip\r |
| 527 | \r |
| 528 | tst r9, #0x080000 @ hflip?\r |
| 529 | bne .rtr_hflip\r |
| 530 | \r |
| 531 | @ Tile (r1=pdest, r3=pal, r9=prevcode, r10=Pico.vram) r2,r4,r7: scratch, r0=0xf\r |
| 532 | Tile 0, 0\r |
| 533 | b .rtrloop\r |
| 534 | \r |
| 535 | .rtr_hflip:\r |
| 536 | Tile 1, 0\r |
| 537 | b .rtrloop\r |
| 538 | \r |
| 539 | .rtr_vflip:\r |
| 540 | tst r9, #0x080000 @ hflip?\r |
| 541 | bne .rtr_vflip_hflip\r |
| 542 | \r |
| 543 | Tile 0, 1\r |
| 544 | b .rtrloop\r |
| 545 | \r |
| 546 | .rtr_vflip_hflip:\r |
| 547 | Tile 1, 1\r |
| 548 | b .rtrloop\r |
| 549 | \r |
| 550 | .rtr_singlecolor:\r |
| 551 | TileSinglecol\r |
| 552 | b .rtrloop\r |
| 553 | \r |
| 554 | .rtr_hiprio:\r |
| 555 | @ *(*hcache)++ = code|(dx<<16)|(trow<<27);\r |
| 556 | sub r4, r1, r11\r |
| 557 | orr r7, r7, r4, lsl #16\r |
| 558 | and r4, r5, #0x00ff0000\r |
| 559 | orr r7, r7, r4, lsl #11 @ (trow<<27)\r |
| 560 | str r7, [r6], #4 @ cache hi priority tile\r |
| 561 | b .rtrloop\r |
| 562 | \r |
| 563 | .rtrloop_exit:\r |
| 564 | add r5, r5, #0x00010000\r |
| 565 | mov r4, r5, lsl #8\r |
| 566 | cmp r4, lr, lsl #24\r |
| 567 | bge .rtrloop_outer_exit\r |
| 568 | add r11, r11, #328*8\r |
| 569 | b .rtrloop_outer\r |
| 570 | \r |
| 571 | .rtrloop_outer_exit:\r |
| 572 | \r |
| 573 | @ terminate cache list\r |
| 574 | mov r0, #0\r |
| 575 | str r0, [r6] @ save cache pointer\r |
| 576 | \r |
| 577 | ldmfd sp!, {r4-r11,lr}\r |
| 578 | bx lr\r |
| 579 | \r |
| 580 | .pool\r |
| 581 | \r |
| 582 | \r |
| 583 | @ void DrawTilesFromCacheF(int *hc, struct PicoEState *est)\r |
| 584 | \r |
| 585 | .global DrawTilesFromCacheF\r |
| 586 | \r |
| 587 | DrawTilesFromCacheF:\r |
| 588 | stmfd sp!, {r4-r10,lr}\r |
| 589 | \r |
| 590 | mov r9, #0xff000000 @ r9=prevcode=-1\r |
| 591 | mvn r6, #0 @ r6=prevy=-1\r |
| 592 | \r |
| 593 | ldr r4, [r1, #OFS_EST_Draw2FB]\r |
| 594 | ldr r2, [r0], #4 @ read y offset\r |
| 595 | mov r7, #328\r |
| 596 | mla r2, r7, r2, r4\r |
| 597 | sub r12, r2, #(328*8*START_ROW) @ r12=scrpos\r |
| 598 | \r |
| 599 | ldr r10, [r1, #OFS_EST_PicoMem_vram]\r |
| 600 | mov r8, r0 @ hc\r |
| 601 | mov r0, #0xf\r |
| 602 | \r |
| 603 | @ scratch: r4, r7\r |
| 604 | @ *hcache++ = code|(dx<<16)|(trow<<27); // cache it\r |
| 605 | \r |
| 606 | .dtfcf_loop:\r |
| 607 | ldr r7, [r8], #4 @ read code\r |
| 608 | movs r1, r7, lsr #16 @ r1=dx;\r |
| 609 | ldmeqfd sp!, {r4-r10,pc} @ dx is never zero, this must be a terminator, return\r |
| 610 | \r |
| 611 | @ row changed?\r |
| 612 | cmp r6, r7, lsr #27\r |
| 613 | movne r6, r7, lsr #27\r |
| 614 | movne r4, #328*8\r |
| 615 | mlane r5, r4, r6, r12 @ r5=pd = scrpos + prevy*328*8\r |
| 616 | \r |
| 617 | bic r1, r1, #0xf800\r |
| 618 | add r1, r5, r1 @ r1=pdest (halfwords)\r |
| 619 | \r |
| 620 | mov r7, r7, lsl #16\r |
| 621 | mov r7, r7, lsr #16\r |
| 622 | \r |
| 623 | cmp r7, r9, lsr #8\r |
| 624 | bne .dtfcf_notsamecode\r |
| 625 | @ we know stuff about this tile already\r |
| 626 | tst r9, #1\r |
| 627 | bne .dtfcf_loop @ empty tile\r |
| 628 | tst r9, #2\r |
| 629 | bne .dtfcf_singlecolor @ singlecolor tile\r |
| 630 | b .dtfcf_samecode\r |
| 631 | \r |
| 632 | .dtfcf_notsamecode:\r |
| 633 | and r4, r9, #0x600000\r |
| 634 | mov r9, r7, lsl #8 @ remember new code\r |
| 635 | \r |
| 636 | @ update cram val\r |
| 637 | and r7, r7, #0x6000\r |
| 638 | mov r3, r7, asr #9 @ r3=pal=((code&0x6000)>>9);\r |
| 639 | \r |
| 640 | \r |
| 641 | .dtfcf_samecode:\r |
| 642 | \r |
| 643 | tst r9, #0x100000 @ vflip?\r |
| 644 | bne .dtfcf_vflip\r |
| 645 | \r |
| 646 | tst r9, #0x080000 @ hflip?\r |
| 647 | bne .dtfcf_hflip\r |
| 648 | \r |
| 649 | @ Tile (r1=pdest, r3=pal, r9=prevcode, r10=Pico.vram) r2,r4,r7: scratch, r0=0xf\r |
| 650 | Tile 0, 0\r |
| 651 | b .dtfcf_loop\r |
| 652 | \r |
| 653 | .dtfcf_hflip:\r |
| 654 | Tile 1, 0\r |
| 655 | b .dtfcf_loop\r |
| 656 | \r |
| 657 | .dtfcf_vflip:\r |
| 658 | tst r9, #0x080000 @ hflip?\r |
| 659 | bne .dtfcf_vflip_hflip\r |
| 660 | \r |
| 661 | Tile 0, 1\r |
| 662 | b .dtfcf_loop\r |
| 663 | \r |
| 664 | .dtfcf_vflip_hflip:\r |
| 665 | Tile 1, 1\r |
| 666 | b .dtfcf_loop\r |
| 667 | \r |
| 668 | .dtfcf_singlecolor:\r |
| 669 | TileSinglecol\r |
| 670 | b .dtfcf_loop\r |
| 671 | \r |
| 672 | .pool\r |
| 673 | \r |
| 674 | \r |
| 675 | @ @@@@@@@@@@@@@@@\r |
| 676 | \r |
| 677 | @ (tile_start<<16)|row_start\r |
| 678 | @ void DrawWindowFull(int start, int end, int prio, struct PicoEState *est)\r |
| 679 | \r |
| 680 | .global DrawWindowFull\r |
| 681 | \r |
| 682 | DrawWindowFull:\r |
| 683 | stmfd sp!, {r4-r11,lr}\r |
| 684 | \r |
| 685 | ldr r11, [r3, #OFS_EST_Pico]\r |
| 686 | ldrb r12, [r11, #OFS_Pico_video_reg+3] @ pvid->reg[3]\r |
| 687 | mov r12, r12, lsl #10\r |
| 688 | \r |
| 689 | ldr r4, [r11, #OFS_Pico_video_reg+12]\r |
| 690 | mov r5, #1 @ nametab_step\r |
| 691 | tst r4, #1 @ 40 cell mode?\r |
| 692 | andne r12, r12, #0xf000 @ 0x3c<<10\r |
| 693 | andeq r12, r12, #0xf800\r |
| 694 | movne r5, r5, lsl #7\r |
| 695 | moveq r5, r5, lsl #6 @ nametab_step\r |
| 696 | \r |
| 697 | and r4, r0, #0xff\r |
| 698 | mla r12, r5, r4, r12 @ nametab += nametab_step*start;\r |
| 699 | \r |
| 700 | ldr r10, [r3, #OFS_EST_PicoMem_vram]\r |
| 701 | mov r4, r0, lsr #16 @ r4=start_cell_h\r |
| 702 | add r7, r12, r4, lsl #1\r |
| 703 | \r |
| 704 | @ fetch the first code now\r |
| 705 | ldrh r7, [r10, r7]\r |
| 706 | cmp r2, r7, lsr #15\r |
| 707 | ldmnefd sp!, {r4-r11,pc} @ hack: simply assume that whole window uses same priority\r |
| 708 | \r |
| 709 | rsb r8, r4, r1, lsr #16 @ cells (h)\r |
| 710 | orr r8, r8, r4, lsl #8\r |
| 711 | mov r4, r1, lsl #24\r |
| 712 | sub r4, r4, r0, lsl #24\r |
| 713 | orr r8, r8, r4, lsr #8 @ r8=cells_h|(start_cell_h<<8)|(cells_v<<16)\r |
| 714 | sub r8, r8, #0x010000 @ adjust for algo\r |
| 715 | \r |
| 716 | mov r9, #0xff000000 @ r9=prevcode=-1\r |
| 717 | \r |
| 718 | ldr r11, [r3, #OFS_EST_Draw2FB]\r |
| 719 | and r4, r0, #0xff\r |
| 720 | add r11, r11, #328*8\r |
| 721 | sub r4, r4, #START_ROW\r |
| 722 | add r11, r11, #8\r |
| 723 | \r |
| 724 | mov r7, #328*8\r |
| 725 | mla r11, r7, r4, r11 @ scrpos+=8*328*(start-START_ROW);\r |
| 726 | mov r0, #0xf\r |
| 727 | \r |
| 728 | .dwfloop_outer:\r |
| 729 | and r6, r8, #0xff00 @ r6=tilex\r |
| 730 | add r1, r11, r6, lsr #5 @ r1=pdest\r |
| 731 | add r6, r12, r6, lsr #7\r |
| 732 | add r6, r10, r6 @ r6=Pico.vram+nametab+tilex\r |
| 733 | orr r8, r8, r8, lsl #24\r |
| 734 | sub r8, r8, #0x01000000 @ cell loop counter\r |
| 735 | b .dwfloop_enter\r |
| 736 | \r |
| 737 | @ r4 & r7 are scratch in this loop\r |
| 738 | .dwfloop:\r |
| 739 | add r1, r1, #8\r |
| 740 | subs r8, r8, #0x01000000\r |
| 741 | bmi .dwfloop_exit\r |
| 742 | \r |
| 743 | .dwfloop_enter:\r |
| 744 | ldrh r7, [r6], #2 @ r7=code\r |
| 745 | \r |
| 746 | cmp r7, r9, lsr #8\r |
| 747 | bne .dwf_notsamecode\r |
| 748 | @ we know stuff about this tile already\r |
| 749 | tst r9, #1\r |
| 750 | bne .dwfloop @ empty tile\r |
| 751 | tst r9, #2\r |
| 752 | bne .dwf_singlecolor @ singlecolor tile\r |
| 753 | b .dwf_samecode\r |
| 754 | \r |
| 755 | .dwf_notsamecode:\r |
| 756 | and r4, r9, #0x600000\r |
| 757 | mov r9, r7, lsl #8 @ remember new code\r |
| 758 | \r |
| 759 | @ update cram val\r |
| 760 | and r7, r7, #0x6000\r |
| 761 | mov r3, r7, asr #9 @ r3=pal=((code&0x6000)>>9);\r |
| 762 | \r |
| 763 | .dwf_samecode:\r |
| 764 | \r |
| 765 | tst r9, #0x100000 @ vflip?\r |
| 766 | bne .dwf_vflip\r |
| 767 | \r |
| 768 | tst r9, #0x080000 @ hflip?\r |
| 769 | bne .dwf_hflip\r |
| 770 | \r |
| 771 | @ Tile (r1=pdest, r3=pal, r9=prevcode, r10=PicoMem.vram)\r |
| 772 | @ r2,r4,r7: scratch, r0=0xf\r |
| 773 | Tile 0, 0\r |
| 774 | b .dwfloop\r |
| 775 | \r |
| 776 | .dwf_hflip:\r |
| 777 | Tile 1, 0\r |
| 778 | b .dwfloop\r |
| 779 | \r |
| 780 | .dwf_vflip:\r |
| 781 | tst r9, #0x080000 @ hflip?\r |
| 782 | bne .dwf_vflip_hflip\r |
| 783 | \r |
| 784 | Tile 0, 1\r |
| 785 | b .dwfloop\r |
| 786 | \r |
| 787 | .dwf_vflip_hflip:\r |
| 788 | Tile 1, 1\r |
| 789 | b .dwfloop\r |
| 790 | \r |
| 791 | .dwf_singlecolor:\r |
| 792 | TileSinglecol 1\r |
| 793 | b .dwfloop\r |
| 794 | \r |
| 795 | .dwfloop_exit:\r |
| 796 | bic r8, r8, #0xff000000 @ fix r8\r |
| 797 | subs r8, r8, #0x010000\r |
| 798 | ldmmifd sp!, {r4-r11,pc}\r |
| 799 | add r11, r11, #328*8\r |
| 800 | add r12, r12, r5 @ nametab+=nametab_step\r |
| 801 | b .dwfloop_outer\r |
| 802 | \r |
| 803 | .pool\r |
| 804 | \r |
| 805 | \r |
| 806 | @ ---------------- sprites ---------------\r |
| 807 | \r |
| 808 | .macro SpriteLoop hflip vflip\r |
| 809 | .if \vflip\r |
| 810 | mov r1, r5, lsr #24 @ height\r |
| 811 | mov r0, #328*8\r |
| 812 | mla r11, r1, r0, r11 @ scrpos+=height*328*8;\r |
| 813 | add r12, r12, r1, lsl #3 @ sy+=height*8\r |
| 814 | .endif\r |
| 815 | mov r0, #0xf\r |
| 816 | .if \hflip\r |
| 817 | and r1, r5, #0xff\r |
| 818 | add r8, r8, r1, lsl #3 @ sx+=width*8\r |
| 819 | 58:\r |
| 820 | cmp r8, #336\r |
| 821 | blt 51f\r |
| 822 | add r9, r9, r5, lsr #16\r |
| 823 | sub r5, r5, #1 @ sub width\r |
| 824 | sub r8, r8, #8\r |
| 825 | b 58b\r |
| 826 | .else\r |
| 827 | cmp r8, #0 @ skip tiles hidden on the left of screen\r |
| 828 | bgt 51f\r |
| 829 | 58:\r |
| 830 | add r9, r9, r5, lsr #16\r |
| 831 | sub r5, r5, #1\r |
| 832 | adds r8, r8, #8\r |
| 833 | ble 58b\r |
| 834 | b 51f\r |
| 835 | .endif\r |
| 836 | \r |
| 837 | 50: @ outer\r |
| 838 | .if !\hflip\r |
| 839 | add r8, r8, #8 @ sx+=8\r |
| 840 | .endif\r |
| 841 | bic r5, r5, #0xff000000 @ fix height\r |
| 842 | orr r5, r5, r5, lsl #16\r |
| 843 | \r |
| 844 | 51: @ outer_enter\r |
| 845 | sub r5, r5, #1 @ width--\r |
| 846 | movs r1, r5, lsl #24\r |
| 847 | ldmmifd sp!, {r4-r11,pc} @ end of tile\r |
| 848 | .if \hflip\r |
| 849 | subs r8, r8, #8 @ sx-=8\r |
| 850 | ldmlefd sp!, {r4-r11,pc} @ tile offscreen\r |
| 851 | .else\r |
| 852 | cmp r8, #328\r |
| 853 | ldmgefd sp!, {r4-r11,pc} @ tile offscreen\r |
| 854 | .endif\r |
| 855 | mov r6, r12 @ r6=sy\r |
| 856 | add r1, r11, r8 @ pdest=scrpos+sx\r |
| 857 | b 53f\r |
| 858 | \r |
| 859 | 52: @ inner\r |
| 860 | add r9, r9, #1<<8 @ tile++\r |
| 861 | .if !\vflip\r |
| 862 | add r6, r6, #8 @ sy+=8\r |
| 863 | add r1, r1, #328*8\r |
| 864 | .endif\r |
| 865 | \r |
| 866 | 53: @ inner_enter\r |
| 867 | @ end of sprite?\r |
| 868 | subs r5, r5, #0x01000000\r |
| 869 | bmi 50b @ ->outer\r |
| 870 | .if \vflip\r |
| 871 | sub r6, r6, #8 @ sy-=8\r |
| 872 | sub r1, r1, #328*8\r |
| 873 | .endif\r |
| 874 | \r |
| 875 | @ offscreen?\r |
| 876 | cmp r6, #(START_ROW*8)\r |
| 877 | ble 52b\r |
| 878 | \r |
| 879 | cmp r6, #(END_ROW*8+8)\r |
| 880 | bge 52b\r |
| 881 | \r |
| 882 | @ Tile (r1=pdest, r3=pal, r9=prevcode, r10=PicoMem.vram)\r |
| 883 | @ r2,r4,r7: scratch, r0=0xf\r |
| 884 | Tile \hflip, \vflip\r |
| 885 | b 52b\r |
| 886 | .endm\r |
| 887 | \r |
| 888 | @ void DrawSpriteFull(unsigned int *sprite, struct PicoEState *est)\r |
| 889 | \r |
| 890 | .global DrawSpriteFull\r |
| 891 | \r |
| 892 | DrawSpriteFull:\r |
| 893 | stmfd sp!, {r4-r11,lr}\r |
| 894 | \r |
| 895 | ldr r3, [r0] @ sprite[0]\r |
| 896 | mov r5, r3, lsl #4\r |
| 897 | mov r6, r5, lsr #30\r |
| 898 | add r6, r6, #1 @ r6=width\r |
| 899 | mov r5, r5, lsl #2\r |
| 900 | mov r5, r5, lsr #30\r |
| 901 | add r5, r5, #1 @ r5=height\r |
| 902 | \r |
| 903 | mov r12, r3, lsl #23\r |
| 904 | mov r12, r12, lsr #23\r |
| 905 | \r |
| 906 | ldr lr, [r0, #4] @ lr=code\r |
| 907 | sub r12, r12, #0x78 @ r12=sy\r |
| 908 | mov r8, lr, lsl #7\r |
| 909 | mov r8, r8, lsr #23\r |
| 910 | sub r8, r8, #0x78 @ r8=sx\r |
| 911 | \r |
| 912 | mov r9, lr, lsl #21\r |
| 913 | mov r9, r9, lsr #13 @ r9=tile<<8\r |
| 914 | \r |
| 915 | and r3, lr, #0x6000\r |
| 916 | mov r3, r3, lsr #9 @ r3=pal=((code>>9)&0x30);\r |
| 917 | \r |
| 918 | ldr r11, [r1, #OFS_EST_Draw2FB]\r |
| 919 | ldr r10, [r1, #OFS_EST_PicoMem_vram]\r |
| 920 | sub r1, r12, #(START_ROW*8)\r |
| 921 | mov r0, #328\r |
| 922 | mla r11, r1, r0, r11 @ scrpos+=(sy-START_ROW*8)*328;\r |
| 923 | \r |
| 924 | orr r5, r5, r5, lsl #16 @\r |
| 925 | orr r5, r6, r5, lsl #8 @ r5=width|(height<<8)|(height<<24)\r |
| 926 | \r |
| 927 | tst lr, #0x1000 @ vflip?\r |
| 928 | bne .dsf_vflip\r |
| 929 | \r |
| 930 | tst lr, #0x0800 @ hflip?\r |
| 931 | bne .dsf_hflip\r |
| 932 | \r |
| 933 | SpriteLoop 0, 0\r |
| 934 | \r |
| 935 | .dsf_hflip:\r |
| 936 | SpriteLoop 1, 0\r |
| 937 | \r |
| 938 | .dsf_vflip:\r |
| 939 | tst lr, #0x0800 @ hflip?\r |
| 940 | bne .dsf_vflip_hflip\r |
| 941 | \r |
| 942 | SpriteLoop 0, 1\r |
| 943 | \r |
| 944 | .dsf_vflip_hflip:\r |
| 945 | SpriteLoop 1, 1\r |
| 946 | \r |
| 947 | .pool\r |
| 948 | \r |
| 949 | @ vim:filetype=armasm\r |