cff531af |
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 |
cc68a136 |
10 | \r |
7b4813fa |
11 | #include "pico_int_o32.h"\r |
cc68a136 |
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 |
f62850ba |
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 |
cc68a136 |
26 | \r |
7b4813fa |
27 | @ void BackFillFull(void *dst, int reg7)\r |
28 | \r |
29 | .global BackFillFull\r |
cc68a136 |
30 | \r |
31 | BackFillFull:\r |
32 | stmfd sp!, {r4-r9,lr}\r |
33 | \r |
7b4813fa |
34 | add lr, r0, #328*8\r |
35 | mov r0, r1, lsl #26\r |
ea8c405f |
36 | mov r0, r0, lsr #26\r |
cc68a136 |
37 | \r |
cc68a136 |
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 |
7b4813fa |
69 | ldmfd sp!, {r4-r9,lr}\r |
70 | bx lr\r |
cc68a136 |
71 | \r |
72 | .pool\r |
73 | \r |
74 | @ -------- some macros --------\r |
75 | \r |
76 | \r |
77 | @ helper\r |
78 | @ TileLineSinglecol (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r0: pixels8_old\r |
79 | .macro TileLineSinglecol notsinglecol=0\r |
80 | and r2, r2, #0xf @ #0x0000000f\r |
81 | .if !\notsinglecol\r |
82 | cmp r2, r0, lsr #28 @ if these don't match,\r |
83 | bicne r9, r9, #2 @ it is a sign that whole tile is not singlecolor (only it's lines may be)\r |
84 | .endif\r |
85 | orr r4, r3, r2\r |
86 | orr r4, r4, r4, lsl #8\r |
87 | \r |
88 | tst r1, #1 @ not aligned?\r |
89 | strneb r4, [r1], #1\r |
90 | streqh r4, [r1], #2\r |
91 | strh r4, [r1], #2\r |
92 | strh r4, [r1], #2\r |
93 | strh r4, [r1], #2\r |
94 | strneb r4, [r1], #1 @ have a remaining unaligned pixel?\r |
95 | sub r1, r1, #8\r |
96 | .if !\notsinglecol\r |
97 | mov r0, #0xf\r |
98 | orr r0, r0, r2, lsl #28 @ we will need the old palindex later\r |
99 | .endif\r |
100 | .endm\r |
101 | \r |
102 | @ TileNorm (r1=pdest, r2=pixels8, r3=pal) r0,r4: scratch\r |
103 | .macro TileLineNorm\r |
104 | ands r4, r0, r2, lsr #12 @ #0x0000f000\r |
105 | orrne r4, r3, r4\r |
106 | strneb r4, [r1]\r |
107 | ands r4, r0, r2, lsr #8 @ #0x00000f00\r |
108 | orrne r4, r3, r4\r |
109 | strneb r4, [r1,#1]\r |
110 | ands r4, r0, r2, lsr #4 @ #0x000000f0\r |
111 | orrne r4, r3, r4\r |
112 | strneb r4, [r1,#2]\r |
113 | ands r4, r0, r2 @ #0x0000000f\r |
114 | orrne r4, r3, r4\r |
115 | strneb r4, [r1,#3]\r |
116 | ands r4, r0, r2, lsr #28 @ #0xf0000000\r |
117 | orrne r4, r3, r4\r |
118 | strneb r4, [r1,#4]\r |
119 | ands r4, r0, r2, lsr #24 @ #0x0f000000\r |
120 | orrne r4, r3, r4\r |
121 | strneb r4, [r1,#5]\r |
122 | ands r4, r0, r2, lsr #20 @ #0x00f00000\r |
123 | orrne r4, r3, r4\r |
124 | strneb r4, [r1,#6]\r |
125 | ands r4, r0, r2, lsr #16 @ #0x000f0000\r |
126 | orrne r4, r3, r4\r |
127 | strneb r4, [r1,#7]\r |
128 | .endm\r |
129 | \r |
130 | @ TileFlip (r1=pdest, r2=pixels8, r3=pal) r0,r4: scratch\r |
131 | .macro TileLineFlip\r |
132 | ands r4, r0, r2, lsr #16 @ #0x000f0000\r |
133 | orrne r4, r3, r4\r |
134 | strneb r4, [r1]\r |
135 | ands r4, r0, r2, lsr #20 @ #0x00f00000\r |
136 | orrne r4, r3, r4\r |
137 | strneb r4, [r1,#1]\r |
138 | ands r4, r0, r2, lsr #24 @ #0x0f000000\r |
139 | orrne r4, r3, r4\r |
140 | strneb r4, [r1,#2]\r |
141 | ands r4, r0, r2, lsr #28 @ #0xf0000000\r |
142 | orrne r4, r3, r4\r |
143 | strneb r4, [r1,#3]\r |
144 | ands r4, r0, r2 @ #0x0000000f\r |
145 | orrne r4, r3, r4\r |
146 | strneb r4, [r1,#4]\r |
147 | ands r4, r0, r2, lsr #4 @ #0x000000f0\r |
148 | orrne r4, r3, r4\r |
149 | strneb r4, [r1,#5]\r |
150 | ands r4, r0, r2, lsr #8 @ #0x00000f00\r |
151 | orrne r4, r3, r4\r |
152 | strneb r4, [r1,#6]\r |
153 | ands r4, r0, r2, lsr #12 @ #0x0000f000\r |
154 | orrne r4, r3, r4\r |
155 | strneb r4, [r1,#7]\r |
156 | .endm\r |
157 | \r |
158 | @ Tile (r1=pdest, r3=pal, r9=prevcode, r10=Pico.vram) r2,r4,r7: scratch, r0=0xf\r |
159 | .macro Tile hflip vflip\r |
160 | mov r7, r9, lsl #13 @ r9=code<<8; addr=(code&0x7ff)<<4;\r |
161 | add r7, r10, r7, lsr #16\r |
162 | orr r9, r9, #3 @ emptytile=singlecolor=1, r9 must be <code_16> 00000xxx\r |
163 | .if \vflip\r |
164 | @ we read tilecodes in reverse order if we have vflip\r |
165 | add r7, r7, #8*4\r |
166 | .endif\r |
167 | @ loop through 8 lines\r |
168 | orr r9, r9, #(7<<24)\r |
169 | b 1f @ loop_enter\r |
170 | \r |
171 | 0: @ singlecol_loop\r |
172 | subs r9, r9, #(1<<24)\r |
173 | add r1, r1, #328 @ set pointer to next line\r |
174 | bmi 8f @ loop_exit with r0 restore\r |
175 | 1:\r |
176 | .if \vflip\r |
177 | ldr r2, [r7, #-4]! @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels\r |
178 | .else\r |
179 | ldr r2, [r7], #4\r |
180 | .endif\r |
181 | tst r2, r2\r |
182 | beq 2f @ empty line\r |
183 | bic r9, r9, #1\r |
184 | cmp r2, r2, ror #4\r |
185 | bne 3f @ not singlecolor\r |
186 | TileLineSinglecol\r |
187 | b 0b\r |
188 | \r |
189 | 2:\r |
190 | bic r9, r9, #2\r |
191 | 2: @ empty_loop\r |
192 | subs r9, r9, #(1<<24)\r |
193 | add r1, r1, #328 @ set pointer to next line\r |
194 | bmi 8f @ loop_exit with r0 restore\r |
195 | .if \vflip\r |
196 | ldr r2, [r7, #-4]! @ next pack\r |
197 | .else\r |
198 | ldr r2, [r7], #4\r |
199 | .endif\r |
200 | mov r0, #0xf @ singlecol_loop might have messed r0\r |
201 | tst r2, r2\r |
202 | beq 2b\r |
203 | \r |
204 | bic r9, r9, #3 @ if we are here, it means we have empty and not empty line\r |
205 | b 5f\r |
206 | \r |
207 | 3: @ not empty, not singlecol\r |
208 | mov r0, #0xf\r |
209 | bic r9, r9, #3\r |
210 | b 6f\r |
211 | \r |
212 | 4: @ not empty, not singlecol loop\r |
213 | subs r9, r9, #(1<<24)\r |
214 | add r1, r1, #328 @ set pointer to next line\r |
215 | bmi 9f @ loop_exit\r |
216 | .if \vflip\r |
217 | ldr r2, [r7, #-4]! @ next pack\r |
218 | .else\r |
219 | ldr r2, [r7], #4\r |
220 | .endif\r |
221 | tst r2, r2\r |
222 | beq 4b @ empty line\r |
223 | 5:\r |
224 | cmp r2, r2, ror #4\r |
225 | beq 7f @ singlecolor line\r |
226 | 6:\r |
227 | .if \hflip\r |
228 | TileLineFlip\r |
229 | .else\r |
230 | TileLineNorm\r |
231 | .endif\r |
232 | b 4b\r |
233 | 7:\r |
234 | TileLineSinglecol 1\r |
235 | b 4b\r |
236 | \r |
237 | 8:\r |
238 | mov r0, #0xf\r |
239 | 9: @ loop_exit\r |
240 | add r9, r9, #(1<<24) @ fix r9\r |
241 | sub r1, r1, #328*8 @ restore pdest pointer\r |
242 | .endm\r |
243 | \r |
244 | \r |
245 | @ TileLineSinglecolAl (r1=pdest, r4,r7=color)\r |
246 | .macro TileLineSinglecolAl0\r |
247 | stmia r1!, {r4,r7}\r |
248 | add r1, r1, #320\r |
249 | .endm\r |
250 | \r |
251 | .macro TileLineSinglecolAl1\r |
252 | strb r4, [r1], #1\r |
253 | strh r4, [r1], #2\r |
254 | str r4, [r1], #4\r |
255 | strb r4, [r1], #1+320\r |
256 | @ add r1, r1, #320\r |
257 | .endm\r |
258 | \r |
259 | .macro TileLineSinglecolAl2\r |
260 | strh r4, [r1], #2\r |
261 | str r4, [r1], #4\r |
262 | strh r4, [r1], #2\r |
263 | add r1, r1, #320\r |
264 | .endm\r |
265 | \r |
266 | .macro TileLineSinglecolAl3\r |
267 | strb r4, [r1], #1\r |
268 | str r4, [r1], #4\r |
269 | strh r4, [r1], #2\r |
270 | strb r4, [r1], #1+320\r |
271 | @ add r1, r1, #320\r |
272 | .endm\r |
273 | \r |
274 | @ TileSinglecol (r1=pdest, r2=pixels8, r3=pal) r4,r7: scratch, r0=0xf\r |
275 | @ kaligned==1, if dest is always aligned\r |
276 | .macro TileSinglecol kaligned=0\r |
277 | and r4, r2, #0xf @ we assume we have good r2 from previous time\r |
278 | orr r4, r4, r3\r |
279 | orr r4, r4, r4, lsl #8\r |
280 | orr r4, r4, r4, lsl #16\r |
281 | mov r7, r4\r |
282 | \r |
283 | .if !\kaligned\r |
284 | tst r1, #2 @ not aligned?\r |
285 | bne 2f\r |
286 | tst r1, #1\r |
287 | bne 1f\r |
288 | .endif\r |
289 | \r |
290 | TileLineSinglecolAl0\r |
291 | TileLineSinglecolAl0\r |
292 | TileLineSinglecolAl0\r |
293 | TileLineSinglecolAl0\r |
294 | TileLineSinglecolAl0\r |
295 | TileLineSinglecolAl0\r |
296 | TileLineSinglecolAl0\r |
297 | TileLineSinglecolAl0\r |
298 | \r |
299 | .if !\kaligned\r |
300 | b 4f\r |
301 | 1:\r |
302 | TileLineSinglecolAl1\r |
303 | TileLineSinglecolAl1\r |
304 | TileLineSinglecolAl1\r |
305 | TileLineSinglecolAl1\r |
306 | TileLineSinglecolAl1\r |
307 | TileLineSinglecolAl1\r |
308 | TileLineSinglecolAl1\r |
309 | TileLineSinglecolAl1\r |
310 | b 4f\r |
311 | \r |
312 | 2:\r |
313 | tst r1, #1\r |
314 | bne 3f\r |
315 | \r |
316 | TileLineSinglecolAl2\r |
317 | TileLineSinglecolAl2\r |
318 | TileLineSinglecolAl2\r |
319 | TileLineSinglecolAl2\r |
320 | TileLineSinglecolAl2\r |
321 | TileLineSinglecolAl2\r |
322 | TileLineSinglecolAl2\r |
323 | TileLineSinglecolAl2\r |
324 | b 4f\r |
325 | \r |
326 | 3:\r |
327 | TileLineSinglecolAl3\r |
328 | TileLineSinglecolAl3\r |
329 | TileLineSinglecolAl3\r |
330 | TileLineSinglecolAl3\r |
331 | TileLineSinglecolAl3\r |
332 | TileLineSinglecolAl3\r |
333 | TileLineSinglecolAl3\r |
334 | TileLineSinglecolAl3\r |
335 | \r |
336 | 4:\r |
337 | .endif\r |
338 | sub r1, r1, #328*8 @ restore pdest pointer\r |
339 | .endm\r |
340 | \r |
341 | \r |
342 | \r |
343 | @ DrawLayerTiles(*hcache, *scrpos, (cells<<24)|(nametab<<9)|(vscroll&0x3ff)<<11|(shift[width]<<8)|planeend, (ymask<<24)|(planestart<<16)|[htab||hscroll]\r |
344 | \r |
7b4813fa |
345 | @ void DrawLayerFull(int plane, int *hcache, int planestart, int planeend,\r |
346 | @ struct PicoEState *est)\r |
cc68a136 |
347 | \r |
348 | .global DrawLayerFull\r |
349 | \r |
350 | DrawLayerFull:\r |
7b4813fa |
351 | ldr r12,[sp] @ est\r |
cc68a136 |
352 | stmfd sp!, {r4-r11,lr}\r |
353 | \r |
354 | mov r6, r1 @ hcache\r |
355 | \r |
7b4813fa |
356 | ldr r11, [r12, #OFS_Pico_video]\r |
357 | ldr r10, [r12, #OFS_Pico_vram]\r |
ea8c405f |
358 | ldrb r5, [r11, #13] @ pvid->reg[13]\r |
359 | ldrb r7, [r11, #11]\r |
360 | \r |
cc68a136 |
361 | sub lr, r3, r2\r |
362 | and lr, lr, #0x00ff0000 @ lr=cells\r |
363 | \r |
cc68a136 |
364 | mov r5, r5, lsl #10 @ htab=pvid->reg[13]<<9; (halfwords)\r |
365 | add r5, r5, r0, lsl #1 @ htab+=plane\r |
366 | bic r5, r5, #0x00ff0000 @ just in case\r |
367 | \r |
cc68a136 |
368 | tst r7, #3 @ full screen scroll? (if ==0)\r |
ea8c405f |
369 | ldrb r7, [r11, #16] @ ??hh??ww\r |
cc68a136 |
370 | ldreqh r5, [r10, r5]\r |
371 | biceq r5, r5, #0x0000fc00 @ r5=hscroll (0-0x3ff)\r |
372 | movne r5, r5, lsr #1\r |
373 | orrne r5, r5, #0x8000 @ this marks that we have htab pointer, not hscroll here\r |
374 | \r |
cc68a136 |
375 | and r8, r7, #3\r |
376 | \r |
377 | orr r5, r5, r7, lsl #1+24\r |
378 | orr r5, r5, #0x1f000000\r |
379 | cmp r8, #1\r |
380 | biclt r5, r5, #0x80000000\r |
381 | biceq r5, r5, #0xc0000000\r |
382 | bicgt r5, r5, #0xe0000000\r |
383 | \r |
384 | mov r9, r2, lsl #24\r |
385 | orr r5, r5, r9, lsr #8 @ r5=(ymask<<24)|(trow<<16)|[htab||hscroll]\r |
386 | \r |
387 | add r4, r8, #5\r |
388 | cmp r4, #7\r |
389 | subge r4, r4, #1 @ r4=shift[width] (5,6,6,7)\r |
390 | \r |
391 | orr lr, lr, r4 \r |
392 | orr lr, lr, r3, lsl #24 @ lr=(planeend<<24)|(cells<<16)|shift[width]\r |
393 | \r |
394 | @ calculate xmask:\r |
395 | mov r8, r8, lsl #24+5\r |
396 | orr r8, r8, #0x1f000000\r |
397 | \r |
ea8c405f |
398 | @ Find name table:\r |
cc68a136 |
399 | tst r0, r0\r |
400 | ldreqb r4, [r11, #2]\r |
401 | moveq r4, r4, lsr #3\r |
402 | ldrneb r4, [r11, #4]\r |
403 | and r4, r4, #7\r |
404 | orr lr, lr, r4, lsl #13 @ lr|=nametab_bits{3}<<13\r |
405 | \r |
7b4813fa |
406 | ldr r11,[sp, #9*4] @ est\r |
cc68a136 |
407 | sub r4, r9, #(START_ROW<<24)\r |
7b4813fa |
408 | ldr r11, [r11, #OFS_Draw2FB]\r |
cc68a136 |
409 | mov r4, r4, asr #24\r |
410 | mov r7, #328*8\r |
411 | mla r11, r4, r7, r11 @ scrpos+=8*328*(planestart-START_ROW);\r |
412 | \r |
ea8c405f |
413 | @ Get vertical scroll value:\r |
cc68a136 |
414 | add r7, r10, #0x012000\r |
415 | add r7, r7, #0x000180 @ r7=Pico.vsram (Pico+0x22180)\r |
416 | ldr r7, [r7]\r |
417 | tst r0, r0\r |
418 | moveq r7, r7, lsl #22\r |
419 | movne r7, r7, lsl #6\r |
420 | mov r7, r7, lsr #22 @ r7=vscroll (10 bits)\r |
421 | \r |
422 | orr lr, lr, r7, lsl #3\r |
423 | mov lr, lr, ror #24 @ packed: cccccccc nnnvvvvv vvvvvsss pppppppp: cells, nametab, vscroll, shift[width], planeend\r |
424 | \r |
425 | ands r7, r7, #7\r |
426 | addne lr, lr, #1 @ we have vertically clipped tiles due to vscroll, so we need 1 more row\r |
427 | \r |
428 | rsb r7, r7, #8\r |
429 | str r7, [r6], #4 @ push y-offset to tilecache\r |
430 | mov r4, #328\r |
431 | mla r11, r4, r7, r11 @ scrpos+=(8-(vscroll&7))*328;\r |
432 | \r |
433 | mov r9, #0xff000000 @ r9=(prevcode<<8)|flags: 1~tile empty, 2~tile singlecolor\r |
434 | \r |
435 | .rtrloop_outer:\r |
436 | mov r4, lr, lsl #11\r |
437 | mov r4, r4, lsr #25 @ r4=vscroll>>3 (7 bits)\r |
438 | add r4, r4, r5, lsr #16 @ +trow\r |
439 | and r4, r4, r5, lsr #24 @ &=ymask\r |
440 | mov r7, lr, lsr #8\r |
441 | and r7, r7, #7 @ shift[width]\r |
442 | mov r0, lr, lsr #9\r |
443 | and r0, r0, #0x7000 @ nametab\r |
444 | add r12,r0, r4, lsl r7 @ nametab_row = nametab + (((trow+(vscroll>>3))&ymask)<<shift[width]); \r |
445 | \r |
446 | mov r4, lr, lsr #24\r |
447 | orr r12,r12,r4, lsl #23\r |
448 | mov r12,r12,lsl #1 @ (nametab_row|(cells<<24)) (halfword compliant)\r |
449 | \r |
450 | @ htab?\r |
451 | tst r5, #0x8000\r |
452 | moveq r7, r5, lsl #22 @ hscroll (0-3FFh)\r |
453 | moveq r7, r7, lsr #22\r |
454 | beq .rtr_hscroll_done\r |
455 | \r |
456 | @ get hscroll from htab\r |
457 | mov r7, r5, lsl #17\r |
458 | ands r4, r5, #0x00ff0000\r |
459 | add r7, r7, r4, lsl #5 @ +=trow<<4\r |
460 | andne r4, lr, #0x3800\r |
461 | subne r7, r7, r4, lsl #7 @ if(trow) htaddr-=(vscroll&7)<<1;\r |
462 | mov r7, r7, lsr #16 @ halfwords\r |
463 | ldrh r7, [r10, r7]\r |
464 | \r |
465 | .rtr_hscroll_done:\r |
ea8c405f |
466 | and r8, r8, #0xff000000\r |
cc68a136 |
467 | rsb r4, r7, #0 @ r4=tilex=(-ts->hscroll)>>3\r |
468 | mov r4, r4, asr #3\r |
469 | and r4, r4, #0xff\r |
cc68a136 |
470 | orr r8, r8, r4 @ r8=(xmask<<24)|tilex\r |
471 | \r |
472 | sub r7, r7, #1\r |
473 | and r7, r7, #7\r |
474 | add r7, r7, #1 @ r7=dx=((ts->hscroll-1)&7)+1\r |
475 | \r |
476 | cmp r7, #8\r |
477 | subeq r12,r12, #0x01000000 @ we will loop cells+1 times, so loop less when there is no hscroll\r |
478 | \r |
479 | add r1, r11, r7 @ r1=pdest\r |
480 | mov r0, #0xf\r |
481 | b .rtrloop_enter\r |
482 | \r |
483 | @ r4 & r7 are scratch in this loop\r |
484 | .rtrloop: @ 40-41 times\r |
485 | add r1, r1, #8\r |
486 | subs r12,r12, #0x01000000\r |
487 | add r8, r8, #1\r |
488 | bmi .rtrloop_exit\r |
489 | \r |
490 | .rtrloop_enter:\r |
491 | and r7, r8, r8, lsr #24\r |
492 | add r7, r10, r7, lsl #1\r |
493 | bic r4, r12, #0xff000000 @ Pico.vram[nametab_row+(tilex&xmask)];\r |
494 | ldrh r7, [r7, r4] @ r7=code (int, but from unsigned, no sign extend)\r |
495 | \r |
496 | tst r7, #0x8000\r |
497 | bne .rtr_hiprio\r |
498 | \r |
499 | cmp r7, r9, lsr #8\r |
500 | bne .rtr_notsamecode\r |
501 | @ we know stuff about this tile already\r |
502 | tst r9, #1\r |
503 | bne .rtrloop @ empty tile\r |
504 | tst r9, #2\r |
505 | bne .rtr_singlecolor @ singlecolor tile\r |
506 | b .rtr_samecode\r |
507 | \r |
508 | .rtr_notsamecode:\r |
509 | and r4, r9, #0x600000\r |
510 | mov r9, r7, lsl #8 @ remember new code\r |
511 | \r |
512 | @ update cram\r |
513 | and r7, r7, #0x6000\r |
514 | mov r3, r7, asr #9 @ r3=pal=((code&0x6000)>>9);\r |
515 | \r |
516 | .rtr_samecode:\r |
517 | tst r9, #0x100000 @ vflip?\r |
518 | bne .rtr_vflip\r |
519 | \r |
520 | tst r9, #0x080000 @ hflip?\r |
521 | bne .rtr_hflip\r |
522 | \r |
523 | @ Tile (r1=pdest, r3=pal, r9=prevcode, r10=Pico.vram) r2,r4,r7: scratch, r0=0xf\r |
524 | Tile 0, 0\r |
525 | b .rtrloop\r |
526 | \r |
527 | .rtr_hflip:\r |
528 | Tile 1, 0\r |
529 | b .rtrloop\r |
530 | \r |
531 | .rtr_vflip:\r |
532 | tst r9, #0x080000 @ hflip?\r |
533 | bne .rtr_vflip_hflip\r |
534 | \r |
535 | Tile 0, 1\r |
536 | b .rtrloop\r |
537 | \r |
538 | .rtr_vflip_hflip:\r |
539 | Tile 1, 1\r |
540 | b .rtrloop\r |
541 | \r |
542 | .rtr_singlecolor:\r |
543 | TileSinglecol\r |
544 | b .rtrloop\r |
545 | \r |
546 | .rtr_hiprio:\r |
547 | @ *(*hcache)++ = code|(dx<<16)|(trow<<27);\r |
548 | sub r4, r1, r11\r |
549 | orr r7, r7, r4, lsl #16\r |
550 | and r4, r5, #0x00ff0000\r |
551 | orr r7, r7, r4, lsl #11 @ (trow<<27)\r |
552 | str r7, [r6], #4 @ cache hi priority tile\r |
553 | b .rtrloop\r |
554 | \r |
555 | .rtrloop_exit:\r |
556 | add r5, r5, #0x00010000\r |
557 | mov r4, r5, lsl #8\r |
558 | cmp r4, lr, lsl #24\r |
559 | bge .rtrloop_outer_exit\r |
560 | add r11, r11, #328*8\r |
561 | b .rtrloop_outer\r |
562 | \r |
563 | .rtrloop_outer_exit:\r |
564 | \r |
565 | @ terminate cache list\r |
566 | mov r0, #0\r |
567 | str r0, [r6] @ save cache pointer\r |
568 | \r |
569 | ldmfd sp!, {r4-r11,lr}\r |
570 | bx lr\r |
571 | \r |
572 | .pool\r |
573 | \r |
574 | \r |
7b4813fa |
575 | @ void DrawTilesFromCacheF(int *hc, struct PicoEState *est)\r |
cc68a136 |
576 | \r |
7b4813fa |
577 | .global DrawTilesFromCacheF\r |
cc68a136 |
578 | \r |
579 | DrawTilesFromCacheF:\r |
580 | stmfd sp!, {r4-r10,lr}\r |
581 | \r |
582 | mov r9, #0xff000000 @ r9=prevcode=-1\r |
583 | mvn r6, #0 @ r6=prevy=-1\r |
584 | \r |
7b4813fa |
585 | ldr r4, [r1, #OFS_Draw2FB]\r |
586 | ldr r2, [r0], #4 @ read y offset\r |
cc68a136 |
587 | mov r7, #328\r |
7b4813fa |
588 | mla r2, r7, r2, r4\r |
589 | sub r12, r2, #(328*8*START_ROW) @ r12=scrpos\r |
cc68a136 |
590 | \r |
7b4813fa |
591 | ldr r10, [r1, #OFS_Pico_vram]\r |
cc68a136 |
592 | mov r8, r0 @ hc\r |
593 | mov r0, #0xf\r |
594 | \r |
595 | @ scratch: r4, r7\r |
596 | @ *hcache++ = code|(dx<<16)|(trow<<27); // cache it\r |
597 | \r |
598 | .dtfcf_loop:\r |
599 | ldr r7, [r8], #4 @ read code\r |
600 | movs r1, r7, lsr #16 @ r1=dx;\r |
601 | ldmeqfd sp!, {r4-r10,pc} @ dx is never zero, this must be a terminator, return\r |
602 | \r |
ea8c405f |
603 | @ row changed?\r |
cc68a136 |
604 | cmp r6, r7, lsr #27\r |
605 | movne r6, r7, lsr #27\r |
606 | movne r4, #328*8\r |
607 | mlane r5, r4, r6, r12 @ r5=pd = scrpos + prevy*328*8\r |
608 | \r |
609 | bic r1, r1, #0xf800\r |
610 | add r1, r5, r1 @ r1=pdest (halfwords)\r |
611 | \r |
612 | mov r7, r7, lsl #16\r |
613 | mov r7, r7, lsr #16\r |
614 | \r |
615 | cmp r7, r9, lsr #8\r |
616 | bne .dtfcf_notsamecode\r |
617 | @ we know stuff about this tile already\r |
618 | tst r9, #1\r |
619 | bne .dtfcf_loop @ empty tile\r |
620 | tst r9, #2\r |
621 | bne .dtfcf_singlecolor @ singlecolor tile\r |
622 | b .dtfcf_samecode\r |
623 | \r |
624 | .dtfcf_notsamecode:\r |
625 | and r4, r9, #0x600000\r |
626 | mov r9, r7, lsl #8 @ remember new code\r |
627 | \r |
628 | @ update cram val\r |
629 | and r7, r7, #0x6000\r |
630 | mov r3, r7, asr #9 @ r3=pal=((code&0x6000)>>9);\r |
631 | \r |
632 | \r |
633 | .dtfcf_samecode:\r |
634 | \r |
635 | tst r9, #0x100000 @ vflip?\r |
636 | bne .dtfcf_vflip\r |
637 | \r |
638 | tst r9, #0x080000 @ hflip?\r |
639 | bne .dtfcf_hflip\r |
640 | \r |
641 | @ Tile (r1=pdest, r3=pal, r9=prevcode, r10=Pico.vram) r2,r4,r7: scratch, r0=0xf\r |
642 | Tile 0, 0\r |
643 | b .dtfcf_loop\r |
644 | \r |
645 | .dtfcf_hflip:\r |
646 | Tile 1, 0\r |
647 | b .dtfcf_loop\r |
648 | \r |
649 | .dtfcf_vflip:\r |
650 | tst r9, #0x080000 @ hflip?\r |
651 | bne .dtfcf_vflip_hflip\r |
652 | \r |
653 | Tile 0, 1\r |
654 | b .dtfcf_loop\r |
655 | \r |
656 | .dtfcf_vflip_hflip:\r |
657 | Tile 1, 1\r |
658 | b .dtfcf_loop\r |
659 | \r |
660 | .dtfcf_singlecolor:\r |
661 | TileSinglecol\r |
662 | b .dtfcf_loop\r |
663 | \r |
664 | .pool\r |
665 | \r |
666 | \r |
667 | @ @@@@@@@@@@@@@@@\r |
668 | \r |
669 | @ (tile_start<<16)|row_start\r |
7b4813fa |
670 | @ void DrawWindowFull(int start, int end, int prio, struct PicoEState *est)\r |
671 | \r |
672 | .global DrawWindowFull\r |
cc68a136 |
673 | \r |
674 | DrawWindowFull:\r |
675 | stmfd sp!, {r4-r11,lr}\r |
676 | \r |
7b4813fa |
677 | ldr r11, [r3, #OFS_Pico_video]\r |
cc68a136 |
678 | ldrb r12, [r11, #3] @ pvid->reg[3]\r |
679 | mov r12, r12, lsl #10\r |
680 | \r |
681 | ldr r4, [r11, #12]\r |
682 | mov r5, #1 @ nametab_step\r |
683 | tst r4, #1 @ 40 cell mode?\r |
684 | andne r12, r12, #0xf000 @ 0x3c<<10\r |
685 | andeq r12, r12, #0xf800\r |
686 | movne r5, r5, lsl #7\r |
687 | moveq r5, r5, lsl #6 @ nametab_step\r |
688 | \r |
689 | and r4, r0, #0xff\r |
690 | mla r12, r5, r4, r12 @ nametab += nametab_step*start;\r |
691 | \r |
7b4813fa |
692 | ldr r10, [r3, #OFS_Pico_vram]\r |
cc68a136 |
693 | mov r4, r0, lsr #16 @ r4=start_cell_h\r |
694 | add r7, r12, r4, lsl #1\r |
695 | \r |
696 | @ fetch the first code now\r |
cc68a136 |
697 | ldrh r7, [r10, r7]\r |
698 | cmp r2, r7, lsr #15\r |
699 | ldmnefd sp!, {r4-r11,pc} @ hack: simply assume that whole window uses same priority\r |
700 | \r |
701 | rsb r8, r4, r1, lsr #16 @ cells (h)\r |
702 | orr r8, r8, r4, lsl #8\r |
703 | mov r4, r1, lsl #24\r |
704 | sub r4, r4, r0, lsl #24\r |
705 | orr r8, r8, r4, lsr #8 @ r8=cells_h|(start_cell_h<<8)|(cells_v<<16)\r |
706 | sub r8, r8, #0x010000 @ adjust for algo\r |
707 | \r |
708 | mov r9, #0xff000000 @ r9=prevcode=-1\r |
709 | \r |
7b4813fa |
710 | ldr r11, [r3, #OFS_Draw2FB]\r |
ea8c405f |
711 | and r4, r0, #0xff\r |
cc68a136 |
712 | add r11, r11, #328*8\r |
7b4813fa |
713 | sub r4, r4, #START_ROW\r |
cc68a136 |
714 | add r11, r11, #8\r |
715 | \r |
cc68a136 |
716 | mov r7, #328*8\r |
717 | mla r11, r7, r4, r11 @ scrpos+=8*328*(start-START_ROW);\r |
718 | mov r0, #0xf\r |
719 | \r |
720 | .dwfloop_outer:\r |
721 | and r6, r8, #0xff00 @ r6=tilex\r |
722 | add r1, r11, r6, lsr #5 @ r1=pdest\r |
723 | add r6, r12, r6, lsr #7\r |
724 | add r6, r10, r6 @ r6=Pico.vram+nametab+tilex\r |
725 | orr r8, r8, r8, lsl #24\r |
726 | sub r8, r8, #0x01000000 @ cell loop counter\r |
727 | b .dwfloop_enter\r |
728 | \r |
729 | @ r4 & r7 are scratch in this loop\r |
730 | .dwfloop:\r |
731 | add r1, r1, #8\r |
732 | subs r8, r8, #0x01000000\r |
733 | bmi .dwfloop_exit\r |
734 | \r |
735 | .dwfloop_enter:\r |
736 | ldrh r7, [r6], #2 @ r7=code\r |
737 | \r |
738 | cmp r7, r9, lsr #8\r |
739 | bne .dwf_notsamecode\r |
740 | @ we know stuff about this tile already\r |
741 | tst r9, #1\r |
742 | bne .dwfloop @ empty tile\r |
743 | tst r9, #2\r |
744 | bne .dwf_singlecolor @ singlecolor tile\r |
745 | b .dwf_samecode\r |
746 | \r |
747 | .dwf_notsamecode:\r |
748 | and r4, r9, #0x600000\r |
749 | mov r9, r7, lsl #8 @ remember new code\r |
750 | \r |
751 | @ update cram val\r |
752 | and r7, r7, #0x6000\r |
753 | mov r3, r7, asr #9 @ r3=pal=((code&0x6000)>>9);\r |
754 | \r |
755 | .dwf_samecode:\r |
756 | \r |
757 | tst r9, #0x100000 @ vflip?\r |
758 | bne .dwf_vflip\r |
759 | \r |
760 | tst r9, #0x080000 @ hflip?\r |
761 | bne .dwf_hflip\r |
762 | \r |
763 | @ Tile (r1=pdest, r3=pal, r9=prevcode, r10=Pico.vram) r2,r4,r7: scratch, r0=0xf\r |
764 | Tile 0, 0\r |
765 | b .dwfloop\r |
766 | \r |
767 | .dwf_hflip:\r |
768 | Tile 1, 0\r |
769 | b .dwfloop\r |
770 | \r |
771 | .dwf_vflip:\r |
772 | tst r9, #0x080000 @ hflip?\r |
773 | bne .dwf_vflip_hflip\r |
774 | \r |
775 | Tile 0, 1\r |
776 | b .dwfloop\r |
777 | \r |
778 | .dwf_vflip_hflip:\r |
779 | Tile 1, 1\r |
780 | b .dwfloop\r |
781 | \r |
782 | .dwf_singlecolor:\r |
783 | TileSinglecol 1\r |
784 | b .dwfloop\r |
785 | \r |
786 | .dwfloop_exit:\r |
787 | bic r8, r8, #0xff000000 @ fix r8\r |
788 | subs r8, r8, #0x010000\r |
789 | ldmmifd sp!, {r4-r11,pc}\r |
790 | add r11, r11, #328*8\r |
791 | add r12, r12, r5 @ nametab+=nametab_step\r |
792 | b .dwfloop_outer\r |
793 | \r |
794 | .pool\r |
795 | \r |
796 | \r |
797 | @ ---------------- sprites ---------------\r |
798 | \r |
799 | .macro SpriteLoop hflip vflip\r |
800 | .if \vflip\r |
801 | mov r1, r5, lsr #24 @ height\r |
802 | mov r0, #328*8\r |
803 | mla r11, r1, r0, r11 @ scrpos+=height*328*8;\r |
804 | add r12, r12, r1, lsl #3 @ sy+=height*8\r |
805 | .endif\r |
806 | mov r0, #0xf\r |
807 | .if \hflip\r |
808 | and r1, r5, #0xff\r |
809 | add r8, r8, r1, lsl #3 @ sx+=width*8\r |
810 | 58:\r |
811 | cmp r8, #336\r |
812 | blt 51f\r |
813 | add r9, r9, r5, lsr #16\r |
814 | sub r5, r5, #1 @ sub width\r |
815 | sub r8, r8, #8\r |
816 | b 58b\r |
817 | .else\r |
818 | cmp r8, #0 @ skip tiles hidden on the left of screen\r |
819 | bgt 51f\r |
820 | 58:\r |
821 | add r9, r9, r5, lsr #16\r |
822 | sub r5, r5, #1\r |
823 | adds r8, r8, #8\r |
824 | ble 58b\r |
825 | b 51f\r |
826 | .endif\r |
827 | \r |
828 | 50: @ outer\r |
829 | .if !\hflip\r |
830 | add r8, r8, #8 @ sx+=8\r |
831 | .endif\r |
832 | bic r5, r5, #0xff000000 @ fix height\r |
833 | orr r5, r5, r5, lsl #16\r |
834 | \r |
835 | 51: @ outer_enter\r |
836 | sub r5, r5, #1 @ width--\r |
837 | movs r1, r5, lsl #24\r |
838 | ldmmifd sp!, {r4-r11,pc} @ end of tile\r |
839 | .if \hflip\r |
840 | subs r8, r8, #8 @ sx-=8\r |
841 | ldmlefd sp!, {r4-r11,pc} @ tile offscreen\r |
842 | .else\r |
843 | cmp r8, #328\r |
844 | ldmgefd sp!, {r4-r11,pc} @ tile offscreen\r |
845 | .endif\r |
846 | mov r6, r12 @ r6=sy\r |
847 | add r1, r11, r8 @ pdest=scrpos+sx\r |
848 | b 53f\r |
849 | \r |
850 | 52: @ inner\r |
851 | add r9, r9, #1<<8 @ tile++\r |
852 | .if !\vflip\r |
853 | add r6, r6, #8 @ sy+=8\r |
854 | add r1, r1, #328*8\r |
855 | .endif\r |
856 | \r |
857 | 53: @ inner_enter\r |
858 | @ end of sprite?\r |
859 | subs r5, r5, #0x01000000\r |
860 | bmi 50b @ ->outer\r |
861 | .if \vflip\r |
862 | sub r6, r6, #8 @ sy-=8\r |
863 | sub r1, r1, #328*8\r |
864 | .endif\r |
865 | \r |
866 | @ offscreen?\r |
867 | cmp r6, #(START_ROW*8)\r |
868 | ble 52b\r |
869 | \r |
870 | cmp r6, #(END_ROW*8+8)\r |
871 | bge 52b\r |
872 | \r |
873 | @ Tile (r1=pdest, r3=pal, r9=prevcode, r10=Pico.vram) r2,r4,r7: scratch, r0=0xf\r |
874 | Tile \hflip, \vflip\r |
875 | b 52b\r |
876 | .endm\r |
877 | \r |
7b4813fa |
878 | @ void DrawSpriteFull(unsigned int *sprite, struct PicoEState *est)\r |
cc68a136 |
879 | \r |
7b4813fa |
880 | .global DrawSpriteFull\r |
cc68a136 |
881 | \r |
882 | DrawSpriteFull:\r |
883 | stmfd sp!, {r4-r11,lr}\r |
884 | \r |
885 | ldr r3, [r0] @ sprite[0]\r |
886 | mov r5, r3, lsl #4\r |
887 | mov r6, r5, lsr #30\r |
888 | add r6, r6, #1 @ r6=width\r |
889 | mov r5, r5, lsl #2\r |
890 | mov r5, r5, lsr #30\r |
891 | add r5, r5, #1 @ r5=height\r |
892 | \r |
893 | mov r12, r3, lsl #23\r |
894 | mov r12, r12, lsr #23\r |
cc68a136 |
895 | \r |
896 | ldr lr, [r0, #4] @ lr=code\r |
ea8c405f |
897 | sub r12, r12, #0x78 @ r12=sy\r |
cc68a136 |
898 | mov r8, lr, lsl #7\r |
899 | mov r8, r8, lsr #23\r |
900 | sub r8, r8, #0x78 @ r8=sx\r |
901 | \r |
902 | mov r9, lr, lsl #21\r |
903 | mov r9, r9, lsr #13 @ r9=tile<<8\r |
904 | \r |
905 | and r3, lr, #0x6000\r |
906 | mov r3, r3, lsr #9 @ r3=pal=((code>>9)&0x30);\r |
907 | \r |
7b4813fa |
908 | ldr r11, [r1, #OFS_Draw2FB]\r |
909 | ldr r10, [r1, #OFS_Pico_vram]\r |
cc68a136 |
910 | sub r1, r12, #(START_ROW*8)\r |
911 | mov r0, #328\r |
912 | mla r11, r1, r0, r11 @ scrpos+=(sy-START_ROW*8)*328;\r |
913 | \r |
914 | orr r5, r5, r5, lsl #16 @\r |
915 | orr r5, r6, r5, lsl #8 @ r5=width|(height<<8)|(height<<24)\r |
916 | \r |
917 | tst lr, #0x1000 @ vflip?\r |
918 | bne .dsf_vflip\r |
919 | \r |
920 | tst lr, #0x0800 @ hflip?\r |
921 | bne .dsf_hflip\r |
922 | \r |
923 | SpriteLoop 0, 0\r |
924 | \r |
925 | .dsf_hflip:\r |
926 | SpriteLoop 1, 0\r |
927 | \r |
928 | .dsf_vflip:\r |
929 | tst lr, #0x0800 @ hflip?\r |
930 | bne .dsf_vflip_hflip\r |
931 | \r |
932 | SpriteLoop 0, 1\r |
933 | \r |
934 | .dsf_vflip_hflip:\r |
935 | SpriteLoop 1, 1\r |
936 | \r |
937 | .pool\r |
938 | \r |
cff531af |
939 | @ vim:filetype=armasm\r |