cff531af |
1 | /*\r |
2 | * assembly optimized versions of most funtions from draw2.c\r |
3 | * (C) notaz, 2006-2008\r |
7bf552b5 |
4 | * (C) irixxxx, 2019-2023\r |
cff531af |
5 | *\r |
6 | * This work is licensed under the terms of MAME license.\r |
7 | * See COPYING file in the top-level directory.\r |
8 | *\r |
9 | * this is highly specialized, be careful if changing related C code!\r |
959ea39b |
10 | *\r |
11 | * NB: this only deals with buffers having line width at 328\r |
cff531af |
12 | */\r |
cc68a136 |
13 | \r |
d40a5af4 |
14 | #include "pico_int_offs.h"\r |
cc68a136 |
15 | \r |
959ea39b |
16 | .equ PDRAW_INTERLACE, (1<<3)\r |
17 | .equ PDRAW_32_COLS, (1<<8)\r |
18 | .equ PDRAW_BORDER_32, (1<<9)\r |
19 | .equ PDRAW_30_ROWS, (1<<11)\r |
20 | \r |
cc68a136 |
21 | @ define these constants in your include file:\r |
22 | @ .equiv START_ROW, 1\r |
23 | @ .equiv END_ROW, 27\r |
24 | @ one row means 8 pixels. If above example was used, (27-1)*8=208 lines would be rendered.\r |
f62850ba |
25 | #ifndef START_ROW\r |
26 | #define START_ROW 0\r |
27 | #endif\r |
28 | #ifndef END_ROW\r |
29 | #define END_ROW 28\r |
30 | #endif\r |
31 | \r |
32 | .text\r |
33 | .align 2\r |
cc68a136 |
34 | \r |
959ea39b |
35 | @ void BackFillFull(unsigned char *dst, int reg7, int lwidth)\r |
98a27142 |
36 | \r |
37 | .global BackFillFull\r |
cc68a136 |
38 | \r |
39 | BackFillFull:\r |
959ea39b |
40 | stmfd sp!, {r4-r10,lr}\r |
41 | \r |
42 | sub r10,r2, #320 @ unused bytes in a line\r |
43 | add lr, r0, #8 @ 8 px overlap area at start of line\r |
44 | add lr, lr, r2, lsl #3 @ 8 lines overlap area at top\r |
cc68a136 |
45 | \r |
98a27142 |
46 | mov r0, r1, lsl #26\r |
ea8c405f |
47 | mov r0, r0, lsr #26\r |
cc68a136 |
48 | \r |
cc68a136 |
49 | orr r0, r0, r0, lsl #8\r |
50 | orr r0, r0, r0, lsl #16\r |
51 | \r |
52 | mov r1, r0 @ 25 opcodes wasted?\r |
53 | mov r2, r0\r |
54 | mov r3, r0\r |
55 | mov r4, r0\r |
56 | mov r5, r0\r |
57 | mov r6, r0\r |
58 | mov r7, r0\r |
59 | mov r8, r0\r |
60 | mov r9, r0\r |
61 | \r |
62 | mov r12, #(END_ROW-START_ROW)*8\r |
63 | \r |
64 | @ go go go!\r |
65 | .bff_loop:\r |
cc68a136 |
66 | subs r12, r12, #1\r |
67 | \r |
68 | stmia lr!, {r0-r9} @ 10*4*8\r |
69 | stmia lr!, {r0-r9}\r |
70 | stmia lr!, {r0-r9}\r |
71 | stmia lr!, {r0-r9}\r |
72 | stmia lr!, {r0-r9}\r |
73 | stmia lr!, {r0-r9}\r |
74 | stmia lr!, {r0-r9}\r |
75 | stmia lr!, {r0-r9}\r |
76 | \r |
959ea39b |
77 | add lr, lr, r10 @ skip unused rest of line\r |
cc68a136 |
78 | bne .bff_loop\r |
79 | \r |
959ea39b |
80 | ldmfd sp!, {r4-r10,lr}\r |
98a27142 |
81 | bx lr\r |
cc68a136 |
82 | \r |
83 | .pool\r |
84 | \r |
85 | @ -------- some macros --------\r |
86 | \r |
cf07a88d |
87 | @ helpers\r |
88 | .macro add_c24 d s c\r |
89 | add \d, \s, #(\c & 0x00ff00)\r |
90 | .if \c & 0x0000ff\r |
91 | add \d, \d, #(\c & 0x0000ff)\r |
92 | .endif\r |
93 | .if \c & 0xff0000\r |
94 | add \d, \d, #(\c & 0xff0000)\r |
95 | .endif\r |
96 | .endm\r |
cc68a136 |
97 | \r |
cc68a136 |
98 | @ TileLineSinglecol (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r0: pixels8_old\r |
99 | .macro TileLineSinglecol notsinglecol=0\r |
100 | and r2, r2, #0xf @ #0x0000000f\r |
101 | .if !\notsinglecol\r |
102 | cmp r2, r0, lsr #28 @ if these don't match,\r |
103 | bicne r9, r9, #2 @ it is a sign that whole tile is not singlecolor (only it's lines may be)\r |
104 | .endif\r |
105 | orr r4, r3, r2\r |
106 | orr r4, r4, r4, lsl #8\r |
107 | \r |
108 | tst r1, #1 @ not aligned?\r |
109 | strneb r4, [r1], #1\r |
110 | streqh r4, [r1], #2\r |
111 | strh r4, [r1], #2\r |
112 | strh r4, [r1], #2\r |
113 | strh r4, [r1], #2\r |
114 | strneb r4, [r1], #1 @ have a remaining unaligned pixel?\r |
115 | sub r1, r1, #8\r |
116 | .if !\notsinglecol\r |
117 | mov r0, #0xf\r |
118 | orr r0, r0, r2, lsl #28 @ we will need the old palindex later\r |
119 | .endif\r |
120 | .endm\r |
121 | \r |
122 | @ TileNorm (r1=pdest, r2=pixels8, r3=pal) r0,r4: scratch\r |
123 | .macro TileLineNorm\r |
124 | ands r4, r0, r2, lsr #12 @ #0x0000f000\r |
125 | orrne r4, r3, r4\r |
126 | strneb r4, [r1]\r |
127 | ands r4, r0, r2, lsr #8 @ #0x00000f00\r |
128 | orrne r4, r3, r4\r |
129 | strneb r4, [r1,#1]\r |
130 | ands r4, r0, r2, lsr #4 @ #0x000000f0\r |
131 | orrne r4, r3, r4\r |
132 | strneb r4, [r1,#2]\r |
133 | ands r4, r0, r2 @ #0x0000000f\r |
134 | orrne r4, r3, r4\r |
135 | strneb r4, [r1,#3]\r |
136 | ands r4, r0, r2, lsr #28 @ #0xf0000000\r |
137 | orrne r4, r3, r4\r |
138 | strneb r4, [r1,#4]\r |
139 | ands r4, r0, r2, lsr #24 @ #0x0f000000\r |
140 | orrne r4, r3, r4\r |
141 | strneb r4, [r1,#5]\r |
142 | ands r4, r0, r2, lsr #20 @ #0x00f00000\r |
143 | orrne r4, r3, r4\r |
144 | strneb r4, [r1,#6]\r |
145 | ands r4, r0, r2, lsr #16 @ #0x000f0000\r |
146 | orrne r4, r3, r4\r |
147 | strneb r4, [r1,#7]\r |
148 | .endm\r |
149 | \r |
150 | @ TileFlip (r1=pdest, r2=pixels8, r3=pal) r0,r4: scratch\r |
151 | .macro TileLineFlip\r |
152 | ands r4, r0, r2, lsr #16 @ #0x000f0000\r |
153 | orrne r4, r3, r4\r |
154 | strneb r4, [r1]\r |
155 | ands r4, r0, r2, lsr #20 @ #0x00f00000\r |
156 | orrne r4, r3, r4\r |
157 | strneb r4, [r1,#1]\r |
158 | ands r4, r0, r2, lsr #24 @ #0x0f000000\r |
159 | orrne r4, r3, r4\r |
160 | strneb r4, [r1,#2]\r |
161 | ands r4, r0, r2, lsr #28 @ #0xf0000000\r |
162 | orrne r4, r3, r4\r |
163 | strneb r4, [r1,#3]\r |
164 | ands r4, r0, r2 @ #0x0000000f\r |
165 | orrne r4, r3, r4\r |
166 | strneb r4, [r1,#4]\r |
167 | ands r4, r0, r2, lsr #4 @ #0x000000f0\r |
168 | orrne r4, r3, r4\r |
169 | strneb r4, [r1,#5]\r |
170 | ands r4, r0, r2, lsr #8 @ #0x00000f00\r |
171 | orrne r4, r3, r4\r |
172 | strneb r4, [r1,#6]\r |
173 | ands r4, r0, r2, lsr #12 @ #0x0000f000\r |
174 | orrne r4, r3, r4\r |
175 | strneb r4, [r1,#7]\r |
176 | .endm\r |
177 | \r |
178 | @ Tile (r1=pdest, r3=pal, r9=prevcode, r10=Pico.vram) r2,r4,r7: scratch, r0=0xf\r |
179 | .macro Tile hflip vflip\r |
180 | mov r7, r9, lsl #13 @ r9=code<<8; addr=(code&0x7ff)<<4;\r |
181 | add r7, r10, r7, lsr #16\r |
182 | orr r9, r9, #3 @ emptytile=singlecolor=1, r9 must be <code_16> 00000xxx\r |
183 | .if \vflip\r |
184 | @ we read tilecodes in reverse order if we have vflip\r |
185 | add r7, r7, #8*4\r |
186 | .endif\r |
187 | @ loop through 8 lines\r |
188 | orr r9, r9, #(7<<24)\r |
189 | b 1f @ loop_enter\r |
190 | \r |
191 | 0: @ singlecol_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 | 1:\r |
196 | .if \vflip\r |
197 | ldr r2, [r7, #-4]! @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels\r |
198 | .else\r |
199 | ldr r2, [r7], #4\r |
200 | .endif\r |
201 | tst r2, r2\r |
202 | beq 2f @ empty line\r |
203 | bic r9, r9, #1\r |
204 | cmp r2, r2, ror #4\r |
205 | bne 3f @ not singlecolor\r |
206 | TileLineSinglecol\r |
207 | b 0b\r |
208 | \r |
209 | 2:\r |
210 | bic r9, r9, #2\r |
211 | 2: @ empty_loop\r |
212 | subs r9, r9, #(1<<24)\r |
213 | add r1, r1, #328 @ set pointer to next line\r |
214 | bmi 8f @ loop_exit with r0 restore\r |
215 | .if \vflip\r |
216 | ldr r2, [r7, #-4]! @ next pack\r |
217 | .else\r |
218 | ldr r2, [r7], #4\r |
219 | .endif\r |
220 | mov r0, #0xf @ singlecol_loop might have messed r0\r |
221 | tst r2, r2\r |
222 | beq 2b\r |
223 | \r |
224 | bic r9, r9, #3 @ if we are here, it means we have empty and not empty line\r |
225 | b 5f\r |
226 | \r |
227 | 3: @ not empty, not singlecol\r |
228 | mov r0, #0xf\r |
229 | bic r9, r9, #3\r |
230 | b 6f\r |
231 | \r |
232 | 4: @ not empty, not singlecol loop\r |
233 | subs r9, r9, #(1<<24)\r |
234 | add r1, r1, #328 @ set pointer to next line\r |
235 | bmi 9f @ loop_exit\r |
236 | .if \vflip\r |
237 | ldr r2, [r7, #-4]! @ next pack\r |
238 | .else\r |
239 | ldr r2, [r7], #4\r |
240 | .endif\r |
241 | tst r2, r2\r |
242 | beq 4b @ empty line\r |
243 | 5:\r |
244 | cmp r2, r2, ror #4\r |
245 | beq 7f @ singlecolor line\r |
246 | 6:\r |
247 | .if \hflip\r |
248 | TileLineFlip\r |
249 | .else\r |
250 | TileLineNorm\r |
251 | .endif\r |
252 | b 4b\r |
253 | 7:\r |
254 | TileLineSinglecol 1\r |
255 | b 4b\r |
256 | \r |
257 | 8:\r |
258 | mov r0, #0xf\r |
259 | 9: @ loop_exit\r |
260 | add r9, r9, #(1<<24) @ fix r9\r |
261 | sub r1, r1, #328*8 @ restore pdest pointer\r |
262 | .endm\r |
263 | \r |
264 | \r |
265 | @ TileLineSinglecolAl (r1=pdest, r4,r7=color)\r |
266 | .macro TileLineSinglecolAl0\r |
267 | stmia r1!, {r4,r7}\r |
268 | add r1, r1, #320\r |
269 | .endm\r |
270 | \r |
271 | .macro TileLineSinglecolAl1\r |
272 | strb r4, [r1], #1\r |
273 | strh r4, [r1], #2\r |
274 | str r4, [r1], #4\r |
275 | strb r4, [r1], #1+320\r |
276 | @ add r1, r1, #320\r |
277 | .endm\r |
278 | \r |
279 | .macro TileLineSinglecolAl2\r |
280 | strh r4, [r1], #2\r |
281 | str r4, [r1], #4\r |
282 | strh r4, [r1], #2\r |
283 | add r1, r1, #320\r |
284 | .endm\r |
285 | \r |
286 | .macro TileLineSinglecolAl3\r |
287 | strb r4, [r1], #1\r |
288 | str r4, [r1], #4\r |
289 | strh r4, [r1], #2\r |
290 | strb r4, [r1], #1+320\r |
291 | @ add r1, r1, #320\r |
292 | .endm\r |
293 | \r |
294 | @ TileSinglecol (r1=pdest, r2=pixels8, r3=pal) r4,r7: scratch, r0=0xf\r |
295 | @ kaligned==1, if dest is always aligned\r |
296 | .macro TileSinglecol kaligned=0\r |
297 | and r4, r2, #0xf @ we assume we have good r2 from previous time\r |
298 | orr r4, r4, r3\r |
299 | orr r4, r4, r4, lsl #8\r |
300 | orr r4, r4, r4, lsl #16\r |
301 | mov r7, r4\r |
302 | \r |
303 | .if !\kaligned\r |
304 | tst r1, #2 @ not aligned?\r |
305 | bne 2f\r |
306 | tst r1, #1\r |
307 | bne 1f\r |
308 | .endif\r |
309 | \r |
310 | TileLineSinglecolAl0\r |
311 | TileLineSinglecolAl0\r |
312 | TileLineSinglecolAl0\r |
313 | TileLineSinglecolAl0\r |
314 | TileLineSinglecolAl0\r |
315 | TileLineSinglecolAl0\r |
316 | TileLineSinglecolAl0\r |
317 | TileLineSinglecolAl0\r |
318 | \r |
319 | .if !\kaligned\r |
320 | b 4f\r |
321 | 1:\r |
322 | TileLineSinglecolAl1\r |
323 | TileLineSinglecolAl1\r |
324 | TileLineSinglecolAl1\r |
325 | TileLineSinglecolAl1\r |
326 | TileLineSinglecolAl1\r |
327 | TileLineSinglecolAl1\r |
328 | TileLineSinglecolAl1\r |
329 | TileLineSinglecolAl1\r |
330 | b 4f\r |
331 | \r |
332 | 2:\r |
333 | tst r1, #1\r |
334 | bne 3f\r |
335 | \r |
336 | TileLineSinglecolAl2\r |
337 | TileLineSinglecolAl2\r |
338 | TileLineSinglecolAl2\r |
339 | TileLineSinglecolAl2\r |
340 | TileLineSinglecolAl2\r |
341 | TileLineSinglecolAl2\r |
342 | TileLineSinglecolAl2\r |
343 | TileLineSinglecolAl2\r |
344 | b 4f\r |
345 | \r |
346 | 3:\r |
347 | TileLineSinglecolAl3\r |
348 | TileLineSinglecolAl3\r |
349 | TileLineSinglecolAl3\r |
350 | TileLineSinglecolAl3\r |
351 | TileLineSinglecolAl3\r |
352 | TileLineSinglecolAl3\r |
353 | TileLineSinglecolAl3\r |
354 | TileLineSinglecolAl3\r |
355 | \r |
356 | 4:\r |
357 | .endif\r |
358 | sub r1, r1, #328*8 @ restore pdest pointer\r |
359 | .endm\r |
360 | \r |
361 | \r |
362 | \r |
363 | @ DrawLayerTiles(*hcache, *scrpos, (cells<<24)|(nametab<<9)|(vscroll&0x3ff)<<11|(shift[width]<<8)|planeend, (ymask<<24)|(planestart<<16)|[htab||hscroll]\r |
364 | \r |
98a27142 |
365 | @ void DrawLayerFull(int plane, int *hcache, int planestart, int planeend,\r |
366 | @ struct PicoEState *est)\r |
cc68a136 |
367 | \r |
368 | .global DrawLayerFull\r |
369 | \r |
370 | DrawLayerFull:\r |
98a27142 |
371 | ldr r12,[sp] @ est\r |
cc68a136 |
372 | stmfd sp!, {r4-r11,lr}\r |
373 | \r |
374 | mov r6, r1 @ hcache\r |
375 | \r |
88fd63ad |
376 | ldr r11, [r12, #OFS_EST_Pico]\r |
377 | ldr r10, [r12, #OFS_EST_PicoMem_vram]\r |
378 | ldrb r5, [r11, #OFS_Pico_video_reg+13] @ pvid->reg[13]\r |
379 | ldrb r7, [r11, #OFS_Pico_video_reg+11]\r |
ea8c405f |
380 | \r |
cc68a136 |
381 | sub lr, r3, r2\r |
382 | and lr, lr, #0x00ff0000 @ lr=cells\r |
383 | \r |
cc68a136 |
384 | mov r5, r5, lsl #10 @ htab=pvid->reg[13]<<9; (halfwords)\r |
385 | add r5, r5, r0, lsl #1 @ htab+=plane\r |
386 | bic r5, r5, #0x00ff0000 @ just in case\r |
387 | \r |
cc68a136 |
388 | tst r7, #3 @ full screen scroll? (if ==0)\r |
88fd63ad |
389 | ldrb r7, [r11, #OFS_Pico_video_reg+16] @ ??hh??ww\r |
cc68a136 |
390 | ldreqh r5, [r10, r5]\r |
391 | biceq r5, r5, #0x0000fc00 @ r5=hscroll (0-0x3ff)\r |
392 | movne r5, r5, lsr #1\r |
393 | orrne r5, r5, #0x8000 @ this marks that we have htab pointer, not hscroll here\r |
394 | \r |
cc68a136 |
395 | and r8, r7, #3\r |
396 | \r |
397 | orr r5, r5, r7, lsl #1+24\r |
398 | orr r5, r5, #0x1f000000\r |
399 | cmp r8, #1\r |
400 | biclt r5, r5, #0x80000000\r |
401 | biceq r5, r5, #0xc0000000\r |
402 | bicgt r5, r5, #0xe0000000\r |
403 | \r |
404 | mov r9, r2, lsl #24\r |
405 | orr r5, r5, r9, lsr #8 @ r5=(ymask<<24)|(trow<<16)|[htab||hscroll]\r |
406 | \r |
407 | add r4, r8, #5\r |
408 | cmp r4, #7\r |
409 | subge r4, r4, #1 @ r4=shift[width] (5,6,6,7)\r |
410 | \r |
411 | orr lr, lr, r4 \r |
412 | orr lr, lr, r3, lsl #24 @ lr=(planeend<<24)|(cells<<16)|shift[width]\r |
413 | \r |
414 | @ calculate xmask:\r |
415 | mov r8, r8, lsl #24+5\r |
416 | orr r8, r8, #0x1f000000\r |
417 | \r |
ea8c405f |
418 | @ Find name table:\r |
cc68a136 |
419 | tst r0, r0\r |
88fd63ad |
420 | ldreqb r4, [r11, #OFS_Pico_video_reg+2]\r |
cc68a136 |
421 | moveq r4, r4, lsr #3\r |
88fd63ad |
422 | ldrneb r4, [r11, #OFS_Pico_video_reg+4]\r |
cc68a136 |
423 | and r4, r4, #7\r |
424 | orr lr, lr, r4, lsl #13 @ lr|=nametab_bits{3}<<13\r |
425 | \r |
98a27142 |
426 | ldr r11,[sp, #9*4] @ est\r |
959ea39b |
427 | ldr r4, [r11, #OFS_EST_Draw2Start]\r |
b9bc876c |
428 | ldr r7, [r11, #OFS_EST_rendstatus]\r |
88fd63ad |
429 | ldr r11, [r11, #OFS_EST_Draw2FB]\r |
959ea39b |
430 | sub r4, r9, r4, lsl #24\r |
431 | tst r7, #PDRAW_BORDER_32 @ H32 border mode?\r |
96a972dc |
432 | tstne r7, #PDRAW_32_COLS\r |
b9bc876c |
433 | addne r11, r11, #32\r |
cc68a136 |
434 | mov r4, r4, asr #24\r |
435 | mov r7, #328*8\r |
959ea39b |
436 | mla r11, r4, r7, r11 @ scrpos+=8*328*(planestart-Draw2Start);\r |
cc68a136 |
437 | \r |
ea8c405f |
438 | @ Get vertical scroll value:\r |
cf07a88d |
439 | add_c24 r7, r10, (OFS_PMEM_vsram-OFS_PMEM_vram)\r |
cc68a136 |
440 | ldr r7, [r7]\r |
441 | tst r0, r0\r |
442 | moveq r7, r7, lsl #22\r |
443 | movne r7, r7, lsl #6\r |
444 | mov r7, r7, lsr #22 @ r7=vscroll (10 bits)\r |
445 | \r |
446 | orr lr, lr, r7, lsl #3\r |
447 | mov lr, lr, ror #24 @ packed: cccccccc nnnvvvvv vvvvvsss pppppppp: cells, nametab, vscroll, shift[width], planeend\r |
448 | \r |
449 | ands r7, r7, #7\r |
450 | addne lr, lr, #1 @ we have vertically clipped tiles due to vscroll, so we need 1 more row\r |
451 | \r |
452 | rsb r7, r7, #8\r |
453 | str r7, [r6], #4 @ push y-offset to tilecache\r |
454 | mov r4, #328\r |
455 | mla r11, r4, r7, r11 @ scrpos+=(8-(vscroll&7))*328;\r |
456 | \r |
457 | mov r9, #0xff000000 @ r9=(prevcode<<8)|flags: 1~tile empty, 2~tile singlecolor\r |
458 | \r |
459 | .rtrloop_outer:\r |
460 | mov r4, lr, lsl #11\r |
461 | mov r4, r4, lsr #25 @ r4=vscroll>>3 (7 bits)\r |
462 | add r4, r4, r5, lsr #16 @ +trow\r |
463 | and r4, r4, r5, lsr #24 @ &=ymask\r |
464 | mov r7, lr, lsr #8\r |
465 | and r7, r7, #7 @ shift[width]\r |
466 | mov r0, lr, lsr #9\r |
467 | and r0, r0, #0x7000 @ nametab\r |
468 | add r12,r0, r4, lsl r7 @ nametab_row = nametab + (((trow+(vscroll>>3))&ymask)<<shift[width]); \r |
469 | \r |
470 | mov r4, lr, lsr #24\r |
471 | orr r12,r12,r4, lsl #23\r |
472 | mov r12,r12,lsl #1 @ (nametab_row|(cells<<24)) (halfword compliant)\r |
473 | \r |
474 | @ htab?\r |
475 | tst r5, #0x8000\r |
476 | moveq r7, r5, lsl #22 @ hscroll (0-3FFh)\r |
477 | moveq r7, r7, lsr #22\r |
478 | beq .rtr_hscroll_done\r |
479 | \r |
480 | @ get hscroll from htab\r |
481 | mov r7, r5, lsl #17\r |
482 | ands r4, r5, #0x00ff0000\r |
483 | add r7, r7, r4, lsl #5 @ +=trow<<4\r |
484 | andne r4, lr, #0x3800\r |
485 | subne r7, r7, r4, lsl #7 @ if(trow) htaddr-=(vscroll&7)<<1;\r |
486 | mov r7, r7, lsr #16 @ halfwords\r |
487 | ldrh r7, [r10, r7]\r |
488 | \r |
489 | .rtr_hscroll_done:\r |
ea8c405f |
490 | and r8, r8, #0xff000000\r |
cc68a136 |
491 | rsb r4, r7, #0 @ r4=tilex=(-ts->hscroll)>>3\r |
492 | mov r4, r4, asr #3\r |
493 | and r4, r4, #0xff\r |
cc68a136 |
494 | orr r8, r8, r4 @ r8=(xmask<<24)|tilex\r |
495 | \r |
496 | sub r7, r7, #1\r |
497 | and r7, r7, #7\r |
498 | add r7, r7, #1 @ r7=dx=((ts->hscroll-1)&7)+1\r |
499 | \r |
500 | cmp r7, #8\r |
501 | subeq r12,r12, #0x01000000 @ we will loop cells+1 times, so loop less when there is no hscroll\r |
502 | \r |
503 | add r1, r11, r7 @ r1=pdest\r |
504 | mov r0, #0xf\r |
505 | b .rtrloop_enter\r |
506 | \r |
507 | @ r4 & r7 are scratch in this loop\r |
508 | .rtrloop: @ 40-41 times\r |
509 | add r1, r1, #8\r |
510 | subs r12,r12, #0x01000000\r |
511 | add r8, r8, #1\r |
512 | bmi .rtrloop_exit\r |
513 | \r |
514 | .rtrloop_enter:\r |
515 | and r7, r8, r8, lsr #24\r |
516 | add r7, r10, r7, lsl #1\r |
517 | bic r4, r12, #0xff000000 @ Pico.vram[nametab_row+(tilex&xmask)];\r |
518 | ldrh r7, [r7, r4] @ r7=code (int, but from unsigned, no sign extend)\r |
519 | \r |
520 | tst r7, #0x8000\r |
521 | bne .rtr_hiprio\r |
522 | \r |
523 | cmp r7, r9, lsr #8\r |
524 | bne .rtr_notsamecode\r |
525 | @ we know stuff about this tile already\r |
526 | tst r9, #1\r |
527 | bne .rtrloop @ empty tile\r |
528 | tst r9, #2\r |
529 | bne .rtr_singlecolor @ singlecolor tile\r |
530 | b .rtr_samecode\r |
531 | \r |
532 | .rtr_notsamecode:\r |
533 | and r4, r9, #0x600000\r |
534 | mov r9, r7, lsl #8 @ remember new code\r |
535 | \r |
536 | @ update cram\r |
537 | and r7, r7, #0x6000\r |
538 | mov r3, r7, asr #9 @ r3=pal=((code&0x6000)>>9);\r |
539 | \r |
540 | .rtr_samecode:\r |
541 | tst r9, #0x100000 @ vflip?\r |
542 | bne .rtr_vflip\r |
543 | \r |
544 | tst r9, #0x080000 @ hflip?\r |
545 | bne .rtr_hflip\r |
546 | \r |
547 | @ Tile (r1=pdest, r3=pal, r9=prevcode, r10=Pico.vram) r2,r4,r7: scratch, r0=0xf\r |
548 | Tile 0, 0\r |
549 | b .rtrloop\r |
550 | \r |
551 | .rtr_hflip:\r |
552 | Tile 1, 0\r |
553 | b .rtrloop\r |
554 | \r |
555 | .rtr_vflip:\r |
556 | tst r9, #0x080000 @ hflip?\r |
557 | bne .rtr_vflip_hflip\r |
558 | \r |
559 | Tile 0, 1\r |
560 | b .rtrloop\r |
561 | \r |
562 | .rtr_vflip_hflip:\r |
563 | Tile 1, 1\r |
564 | b .rtrloop\r |
565 | \r |
566 | .rtr_singlecolor:\r |
567 | TileSinglecol\r |
568 | b .rtrloop\r |
569 | \r |
570 | .rtr_hiprio:\r |
571 | @ *(*hcache)++ = code|(dx<<16)|(trow<<27);\r |
572 | sub r4, r1, r11\r |
573 | orr r7, r7, r4, lsl #16\r |
574 | and r4, r5, #0x00ff0000\r |
575 | orr r7, r7, r4, lsl #11 @ (trow<<27)\r |
576 | str r7, [r6], #4 @ cache hi priority tile\r |
577 | b .rtrloop\r |
578 | \r |
579 | .rtrloop_exit:\r |
580 | add r5, r5, #0x00010000\r |
581 | mov r4, r5, lsl #8\r |
582 | cmp r4, lr, lsl #24\r |
583 | bge .rtrloop_outer_exit\r |
584 | add r11, r11, #328*8\r |
585 | b .rtrloop_outer\r |
586 | \r |
587 | .rtrloop_outer_exit:\r |
588 | \r |
589 | @ terminate cache list\r |
590 | mov r0, #0\r |
591 | str r0, [r6] @ save cache pointer\r |
592 | \r |
593 | ldmfd sp!, {r4-r11,lr}\r |
594 | bx lr\r |
595 | \r |
596 | .pool\r |
597 | \r |
598 | \r |
98a27142 |
599 | @ void DrawTilesFromCacheF(int *hc, struct PicoEState *est)\r |
cc68a136 |
600 | \r |
98a27142 |
601 | .global DrawTilesFromCacheF\r |
cc68a136 |
602 | \r |
603 | DrawTilesFromCacheF:\r |
959ea39b |
604 | stmfd sp!, {r4-r11,lr}\r |
cc68a136 |
605 | \r |
606 | mov r9, #0xff000000 @ r9=prevcode=-1\r |
607 | mvn r6, #0 @ r6=prevy=-1\r |
608 | \r |
b9bc876c |
609 | ldr r7, [r1, #OFS_EST_rendstatus]\r |
88fd63ad |
610 | ldr r4, [r1, #OFS_EST_Draw2FB]\r |
959ea39b |
611 | ldr r11,[r1, #OFS_EST_Draw2Start]\r |
98a27142 |
612 | ldr r2, [r0], #4 @ read y offset\r |
959ea39b |
613 | tst r7, #PDRAW_BORDER_32 @ H32 border mode?\r |
96a972dc |
614 | tstne r7, #PDRAW_32_COLS\r |
b9bc876c |
615 | addne r4, r4, #32\r |
cc68a136 |
616 | mov r7, #328\r |
98a27142 |
617 | mla r2, r7, r2, r4\r |
618 | sub r12, r2, #(328*8*START_ROW) @ r12=scrpos\r |
cc68a136 |
619 | \r |
88fd63ad |
620 | ldr r10, [r1, #OFS_EST_PicoMem_vram]\r |
cc68a136 |
621 | mov r8, r0 @ hc\r |
622 | mov r0, #0xf\r |
623 | \r |
624 | @ scratch: r4, r7\r |
625 | @ *hcache++ = code|(dx<<16)|(trow<<27); // cache it\r |
626 | \r |
627 | .dtfcf_loop:\r |
628 | ldr r7, [r8], #4 @ read code\r |
629 | movs r1, r7, lsr #16 @ r1=dx;\r |
959ea39b |
630 | ldmeqfd sp!, {r4-r11,pc} @ dx is never zero, this must be a terminator, return\r |
cc68a136 |
631 | \r |
ea8c405f |
632 | @ row changed?\r |
cc68a136 |
633 | cmp r6, r7, lsr #27\r |
634 | movne r6, r7, lsr #27\r |
959ea39b |
635 | subne r6, r6, r11\r |
cc68a136 |
636 | movne r4, #328*8\r |
959ea39b |
637 | mlane r5, r4, r6, r12 @ r5=pd = scrpos + (prevy-Draw2Start)*328*8\r |
cc68a136 |
638 | \r |
639 | bic r1, r1, #0xf800\r |
640 | add r1, r5, r1 @ r1=pdest (halfwords)\r |
641 | \r |
642 | mov r7, r7, lsl #16\r |
643 | mov r7, r7, lsr #16\r |
644 | \r |
645 | cmp r7, r9, lsr #8\r |
646 | bne .dtfcf_notsamecode\r |
647 | @ we know stuff about this tile already\r |
648 | tst r9, #1\r |
649 | bne .dtfcf_loop @ empty tile\r |
650 | tst r9, #2\r |
651 | bne .dtfcf_singlecolor @ singlecolor tile\r |
652 | b .dtfcf_samecode\r |
653 | \r |
654 | .dtfcf_notsamecode:\r |
655 | and r4, r9, #0x600000\r |
656 | mov r9, r7, lsl #8 @ remember new code\r |
657 | \r |
658 | @ update cram val\r |
659 | and r7, r7, #0x6000\r |
660 | mov r3, r7, asr #9 @ r3=pal=((code&0x6000)>>9);\r |
661 | \r |
662 | \r |
663 | .dtfcf_samecode:\r |
664 | \r |
665 | tst r9, #0x100000 @ vflip?\r |
666 | bne .dtfcf_vflip\r |
667 | \r |
668 | tst r9, #0x080000 @ hflip?\r |
669 | bne .dtfcf_hflip\r |
670 | \r |
671 | @ Tile (r1=pdest, r3=pal, r9=prevcode, r10=Pico.vram) r2,r4,r7: scratch, r0=0xf\r |
672 | Tile 0, 0\r |
673 | b .dtfcf_loop\r |
674 | \r |
675 | .dtfcf_hflip:\r |
676 | Tile 1, 0\r |
677 | b .dtfcf_loop\r |
678 | \r |
679 | .dtfcf_vflip:\r |
680 | tst r9, #0x080000 @ hflip?\r |
681 | bne .dtfcf_vflip_hflip\r |
682 | \r |
683 | Tile 0, 1\r |
684 | b .dtfcf_loop\r |
685 | \r |
686 | .dtfcf_vflip_hflip:\r |
687 | Tile 1, 1\r |
688 | b .dtfcf_loop\r |
689 | \r |
690 | .dtfcf_singlecolor:\r |
691 | TileSinglecol\r |
692 | b .dtfcf_loop\r |
693 | \r |
694 | .pool\r |
695 | \r |
696 | \r |
697 | @ @@@@@@@@@@@@@@@\r |
698 | \r |
699 | @ (tile_start<<16)|row_start\r |
98a27142 |
700 | @ void DrawWindowFull(int start, int end, int prio, struct PicoEState *est)\r |
701 | \r |
702 | .global DrawWindowFull\r |
cc68a136 |
703 | \r |
704 | DrawWindowFull:\r |
705 | stmfd sp!, {r4-r11,lr}\r |
706 | \r |
88fd63ad |
707 | ldr r11, [r3, #OFS_EST_Pico]\r |
708 | ldrb r12, [r11, #OFS_Pico_video_reg+3] @ pvid->reg[3]\r |
cc68a136 |
709 | mov r12, r12, lsl #10\r |
710 | \r |
88fd63ad |
711 | ldr r4, [r11, #OFS_Pico_video_reg+12]\r |
cc68a136 |
712 | mov r5, #1 @ nametab_step\r |
b9bc876c |
713 | ldr r11, [r3, #OFS_EST_Draw2FB]\r |
959ea39b |
714 | ldr r6, [r3, #OFS_EST_Draw2Start]\r |
cc68a136 |
715 | tst r4, #1 @ 40 cell mode?\r |
716 | andne r12, r12, #0xf000 @ 0x3c<<10\r |
cc68a136 |
717 | movne r5, r5, lsl #7\r |
b9bc876c |
718 | bne 0f\r |
719 | ldr r7, [r3, #OFS_EST_rendstatus]\r |
720 | and r12, r12, #0xf800\r |
721 | mov r5, r5, lsl #6 @ nametab_step\r |
959ea39b |
722 | tst r7, #PDRAW_BORDER_32\r |
96a972dc |
723 | tstne r7, #PDRAW_32_COLS\r |
b9bc876c |
724 | addne r11, r11, #32 @ center screen in H32 mode\r |
725 | \r |
726 | 0: and r4, r0, #0xff\r |
959ea39b |
727 | sub r4, r4, r6\r |
728 | mla r12, r5, r4, r12 @ nametab += nametab_step*(start-Draw2Start];\r |
cc68a136 |
729 | \r |
88fd63ad |
730 | ldr r10, [r3, #OFS_EST_PicoMem_vram]\r |
cc68a136 |
731 | mov r4, r0, lsr #16 @ r4=start_cell_h\r |
732 | add r7, r12, r4, lsl #1\r |
733 | \r |
734 | @ fetch the first code now\r |
cc68a136 |
735 | ldrh r7, [r10, r7]\r |
736 | cmp r2, r7, lsr #15\r |
737 | ldmnefd sp!, {r4-r11,pc} @ hack: simply assume that whole window uses same priority\r |
738 | \r |
739 | rsb r8, r4, r1, lsr #16 @ cells (h)\r |
740 | orr r8, r8, r4, lsl #8\r |
741 | mov r4, r1, lsl #24\r |
742 | sub r4, r4, r0, lsl #24\r |
743 | orr r8, r8, r4, lsr #8 @ r8=cells_h|(start_cell_h<<8)|(cells_v<<16)\r |
744 | sub r8, r8, #0x010000 @ adjust for algo\r |
745 | \r |
746 | mov r9, #0xff000000 @ r9=prevcode=-1\r |
747 | \r |
ea8c405f |
748 | and r4, r0, #0xff\r |
cc68a136 |
749 | add r11, r11, #328*8\r |
959ea39b |
750 | sub r4, r4, r6\r |
cc68a136 |
751 | add r11, r11, #8\r |
752 | \r |
cc68a136 |
753 | mov r7, #328*8\r |
959ea39b |
754 | mla r11, r7, r4, r11 @ scrpos+=8*328*(start-Draw2Start);\r |
cc68a136 |
755 | mov r0, #0xf\r |
756 | \r |
757 | .dwfloop_outer:\r |
758 | and r6, r8, #0xff00 @ r6=tilex\r |
759 | add r1, r11, r6, lsr #5 @ r1=pdest\r |
760 | add r6, r12, r6, lsr #7\r |
761 | add r6, r10, r6 @ r6=Pico.vram+nametab+tilex\r |
762 | orr r8, r8, r8, lsl #24\r |
763 | sub r8, r8, #0x01000000 @ cell loop counter\r |
764 | b .dwfloop_enter\r |
765 | \r |
766 | @ r4 & r7 are scratch in this loop\r |
767 | .dwfloop:\r |
768 | add r1, r1, #8\r |
769 | subs r8, r8, #0x01000000\r |
770 | bmi .dwfloop_exit\r |
771 | \r |
772 | .dwfloop_enter:\r |
773 | ldrh r7, [r6], #2 @ r7=code\r |
774 | \r |
775 | cmp r7, r9, lsr #8\r |
776 | bne .dwf_notsamecode\r |
777 | @ we know stuff about this tile already\r |
778 | tst r9, #1\r |
779 | bne .dwfloop @ empty tile\r |
780 | tst r9, #2\r |
781 | bne .dwf_singlecolor @ singlecolor tile\r |
782 | b .dwf_samecode\r |
783 | \r |
784 | .dwf_notsamecode:\r |
785 | and r4, r9, #0x600000\r |
786 | mov r9, r7, lsl #8 @ remember new code\r |
787 | \r |
788 | @ update cram val\r |
789 | and r7, r7, #0x6000\r |
790 | mov r3, r7, asr #9 @ r3=pal=((code&0x6000)>>9);\r |
791 | \r |
792 | .dwf_samecode:\r |
793 | \r |
794 | tst r9, #0x100000 @ vflip?\r |
795 | bne .dwf_vflip\r |
796 | \r |
797 | tst r9, #0x080000 @ hflip?\r |
798 | bne .dwf_hflip\r |
799 | \r |
88fd63ad |
800 | @ Tile (r1=pdest, r3=pal, r9=prevcode, r10=PicoMem.vram)\r |
801 | @ r2,r4,r7: scratch, r0=0xf\r |
cc68a136 |
802 | Tile 0, 0\r |
803 | b .dwfloop\r |
804 | \r |
805 | .dwf_hflip:\r |
806 | Tile 1, 0\r |
807 | b .dwfloop\r |
808 | \r |
809 | .dwf_vflip:\r |
810 | tst r9, #0x080000 @ hflip?\r |
811 | bne .dwf_vflip_hflip\r |
812 | \r |
813 | Tile 0, 1\r |
814 | b .dwfloop\r |
815 | \r |
816 | .dwf_vflip_hflip:\r |
817 | Tile 1, 1\r |
818 | b .dwfloop\r |
819 | \r |
820 | .dwf_singlecolor:\r |
821 | TileSinglecol 1\r |
822 | b .dwfloop\r |
823 | \r |
824 | .dwfloop_exit:\r |
825 | bic r8, r8, #0xff000000 @ fix r8\r |
826 | subs r8, r8, #0x010000\r |
827 | ldmmifd sp!, {r4-r11,pc}\r |
828 | add r11, r11, #328*8\r |
829 | add r12, r12, r5 @ nametab+=nametab_step\r |
830 | b .dwfloop_outer\r |
831 | \r |
832 | .pool\r |
833 | \r |
834 | \r |
835 | @ ---------------- sprites ---------------\r |
836 | \r |
837 | .macro SpriteLoop hflip vflip\r |
838 | .if \vflip\r |
839 | mov r1, r5, lsr #24 @ height\r |
840 | mov r0, #328*8\r |
841 | mla r11, r1, r0, r11 @ scrpos+=height*328*8;\r |
842 | add r12, r12, r1, lsl #3 @ sy+=height*8\r |
843 | .endif\r |
844 | mov r0, #0xf\r |
845 | .if \hflip\r |
846 | and r1, r5, #0xff\r |
847 | add r8, r8, r1, lsl #3 @ sx+=width*8\r |
848 | 58:\r |
849 | cmp r8, #336\r |
850 | blt 51f\r |
851 | add r9, r9, r5, lsr #16\r |
852 | sub r5, r5, #1 @ sub width\r |
853 | sub r8, r8, #8\r |
854 | b 58b\r |
855 | .else\r |
856 | cmp r8, #0 @ skip tiles hidden on the left of screen\r |
857 | bgt 51f\r |
858 | 58:\r |
859 | add r9, r9, r5, lsr #16\r |
860 | sub r5, r5, #1\r |
861 | adds r8, r8, #8\r |
862 | ble 58b\r |
863 | b 51f\r |
864 | .endif\r |
865 | \r |
866 | 50: @ outer\r |
867 | .if !\hflip\r |
868 | add r8, r8, #8 @ sx+=8\r |
869 | .endif\r |
870 | bic r5, r5, #0xff000000 @ fix height\r |
871 | orr r5, r5, r5, lsl #16\r |
872 | \r |
873 | 51: @ outer_enter\r |
874 | sub r5, r5, #1 @ width--\r |
875 | movs r1, r5, lsl #24\r |
876 | ldmmifd sp!, {r4-r11,pc} @ end of tile\r |
877 | .if \hflip\r |
878 | subs r8, r8, #8 @ sx-=8\r |
879 | ldmlefd sp!, {r4-r11,pc} @ tile offscreen\r |
880 | .else\r |
881 | cmp r8, #328\r |
882 | ldmgefd sp!, {r4-r11,pc} @ tile offscreen\r |
883 | .endif\r |
884 | mov r6, r12 @ r6=sy\r |
885 | add r1, r11, r8 @ pdest=scrpos+sx\r |
886 | b 53f\r |
887 | \r |
888 | 52: @ inner\r |
889 | add r9, r9, #1<<8 @ tile++\r |
890 | .if !\vflip\r |
891 | add r6, r6, #8 @ sy+=8\r |
892 | add r1, r1, #328*8\r |
893 | .endif\r |
894 | \r |
895 | 53: @ inner_enter\r |
896 | @ end of sprite?\r |
897 | subs r5, r5, #0x01000000\r |
898 | bmi 50b @ ->outer\r |
899 | .if \vflip\r |
900 | sub r6, r6, #8 @ sy-=8\r |
901 | sub r1, r1, #328*8\r |
902 | .endif\r |
903 | \r |
904 | @ offscreen?\r |
905 | cmp r6, #(START_ROW*8)\r |
906 | ble 52b\r |
907 | \r |
908 | cmp r6, #(END_ROW*8+8)\r |
909 | bge 52b\r |
910 | \r |
88fd63ad |
911 | @ Tile (r1=pdest, r3=pal, r9=prevcode, r10=PicoMem.vram)\r |
912 | @ r2,r4,r7: scratch, r0=0xf\r |
cc68a136 |
913 | Tile \hflip, \vflip\r |
914 | b 52b\r |
915 | .endm\r |
916 | \r |
98a27142 |
917 | @ void DrawSpriteFull(unsigned int *sprite, struct PicoEState *est)\r |
cc68a136 |
918 | \r |
98a27142 |
919 | .global DrawSpriteFull\r |
cc68a136 |
920 | \r |
921 | DrawSpriteFull:\r |
922 | stmfd sp!, {r4-r11,lr}\r |
923 | \r |
924 | ldr r3, [r0] @ sprite[0]\r |
925 | mov r5, r3, lsl #4\r |
926 | mov r6, r5, lsr #30\r |
927 | add r6, r6, #1 @ r6=width\r |
928 | mov r5, r5, lsl #2\r |
929 | mov r5, r5, lsr #30\r |
930 | add r5, r5, #1 @ r5=height\r |
931 | \r |
932 | mov r12, r3, lsl #23\r |
933 | mov r12, r12, lsr #23\r |
cc68a136 |
934 | \r |
935 | ldr lr, [r0, #4] @ lr=code\r |
ea8c405f |
936 | sub r12, r12, #0x78 @ r12=sy\r |
cc68a136 |
937 | mov r8, lr, lsl #7\r |
938 | mov r8, r8, lsr #23\r |
939 | sub r8, r8, #0x78 @ r8=sx\r |
940 | \r |
941 | mov r9, lr, lsl #21\r |
942 | mov r9, r9, lsr #13 @ r9=tile<<8\r |
943 | \r |
944 | and r3, lr, #0x6000\r |
945 | mov r3, r3, lsr #9 @ r3=pal=((code>>9)&0x30);\r |
946 | \r |
b9bc876c |
947 | ldr r0, [r1, #OFS_EST_rendstatus]\r |
88fd63ad |
948 | ldr r11, [r1, #OFS_EST_Draw2FB]\r |
959ea39b |
949 | ldr r2, [r1, #OFS_EST_Draw2Start]\r |
88fd63ad |
950 | ldr r10, [r1, #OFS_EST_PicoMem_vram]\r |
959ea39b |
951 | tst r0, #PDRAW_BORDER_32 @ H32 border mode?\r |
96a972dc |
952 | tstne r0, #PDRAW_32_COLS\r |
b9bc876c |
953 | addne r11, r11, #32\r |
959ea39b |
954 | sub r12, r12, r2, lsl #3\r |
cc68a136 |
955 | mov r0, #328\r |
959ea39b |
956 | mla r11, r12, r0, r11 @ scrpos+=(sy-Draw2Start*8)*328;\r |
cc68a136 |
957 | \r |
958 | orr r5, r5, r5, lsl #16 @\r |
959 | orr r5, r6, r5, lsl #8 @ r5=width|(height<<8)|(height<<24)\r |
960 | \r |
961 | tst lr, #0x1000 @ vflip?\r |
962 | bne .dsf_vflip\r |
963 | \r |
964 | tst lr, #0x0800 @ hflip?\r |
965 | bne .dsf_hflip\r |
966 | \r |
967 | SpriteLoop 0, 0\r |
968 | \r |
969 | .dsf_hflip:\r |
970 | SpriteLoop 1, 0\r |
971 | \r |
972 | .dsf_vflip:\r |
973 | tst lr, #0x0800 @ hflip?\r |
974 | bne .dsf_vflip_hflip\r |
975 | \r |
976 | SpriteLoop 0, 1\r |
977 | \r |
978 | .dsf_vflip_hflip:\r |
979 | SpriteLoop 1, 1\r |
980 | \r |
981 | .pool\r |
982 | \r |
cff531af |
983 | @ vim:filetype=armasm\r |