2 * assembly optimized versions of most funtions from draw.c
\r
3 * (C) notaz, 2006-2010,2017
\r
5 * This work is licensed under the terms of MAME license.
\r
6 * See COPYING file in the top-level directory.
\r
8 * this is highly specialized, be careful if changing related C code!
\r
11 #include "pico_int_o32.h"
\r
13 .extern DrawStripInterlace
\r
15 .equ PDRAW_SPRITES_MOVED, (1<<0)
\r
16 .equ PDRAW_WND_DIFF_PRIO, (1<<1)
\r
17 .equ PDRAW_ACC_SPRITES, (1<<2)
\r
18 .equ PDRAW_DIRTY_SPRITES, (1<<4)
\r
19 .equ PDRAW_PLANE_HI_PRIO, (1<<6)
\r
20 .equ PDRAW_SHHI_DONE, (1<<7)
\r
23 .macro TilePixel pat lsrr offs
\r
27 ands r4, \pat, r2, lsr #\lsrr
\r
30 strneb r4, [r1,#\offs]
\r
33 @ TileNorm (r1=pdest, r2=pixels8, r3=pal) r4: scratch, pat: register with helper pattern 0xf
\r
35 TilePixel \pat, 12, 0 @ #0x0000f000
\r
36 TilePixel \pat, 8, 1 @ #0x00000f00
\r
37 TilePixel \pat, 4, 2 @ #0x000000f0
\r
38 TilePixel \pat, 0, 3 @ #0x0000000f
\r
39 TilePixel \pat, 28, 4 @ #0xf0000000
\r
40 TilePixel \pat, 24, 5 @ #0x0f000000
\r
41 TilePixel \pat, 20, 6 @ #0x00f00000
\r
42 TilePixel \pat, 16, 7 @ #0x000f0000
\r
45 @ TileFlip (r1=pdest, r2=pixels8, r3=pal) r4: scratch, pat: register with helper pattern 0xf
\r
47 TilePixel \pat, 16, 0 @ #0x000f0000
\r
48 TilePixel \pat, 20, 1 @ #0x00f00000
\r
49 TilePixel \pat, 24, 2 @ #0x0f000000
\r
50 TilePixel \pat, 28, 3 @ #0xf0000000
\r
51 TilePixel \pat, 0, 4 @ #0x0000000f
\r
52 TilePixel \pat, 4, 5 @ #0x000000f0
\r
53 TilePixel \pat, 8, 6 @ #0x00000f00
\r
54 TilePixel \pat, 12, 7 @ #0x0000f000
\r
57 @ shadow/hilight mode
\r
59 @ this one is for hi priority layer
\r
60 .macro TilePixelShHP lsrr offs
\r
64 ands r4, r12, r2, lsr #\lsrr
\r
66 ldreqb r4, [r1,#\offs]
\r
69 strb r4, [r1,#\offs]
\r
72 @ TileNormShHP (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: register with helper pattern 0xf, touches r3 high bits
\r
74 TilePixelShHP 12, 0 @ #0x0000f000
\r
75 TilePixelShHP 8, 1 @ #0x00000f00
\r
76 TilePixelShHP 4, 2 @ #0x000000f0
\r
77 TilePixelShHP 0, 3 @ #0x0000000f
\r
78 TilePixelShHP 28, 4 @ #0xf0000000
\r
79 TilePixelShHP 24, 5 @ #0x0f000000
\r
80 TilePixelShHP 20, 6 @ #0x00f00000
\r
81 TilePixelShHP 16, 7 @ #0x000f0000
\r
84 @ TileFlipShHP (r1=pdest, r2=pixels8, r3=pal) r4: scratch, pat: register with helper pattern 0xf
\r
86 TilePixelShHP 16, 0 @ #0x000f0000
\r
87 TilePixelShHP 20, 1 @ #0x00f00000
\r
88 TilePixelShHP 24, 2 @ #0x0f000000
\r
89 TilePixelShHP 28, 3 @ #0xf0000000
\r
90 TilePixelShHP 0, 4 @ #0x0000000f
\r
91 TilePixelShHP 4, 5 @ #0x000000f0
\r
92 TilePixelShHP 8, 6 @ #0x00000f00
\r
93 TilePixelShHP 12, 7 @ #0x0000f000
\r
97 @ TileSingleSh (r1=pdest, r2=pixels8, r3=pal) r4,r7: scratch, r0=sx; r12: helper pattern 0xf
\r
99 tst r0, #1 @ not aligned?
\r
105 strneb r4, [r1], #1
\r
106 streqh r4, [r1], #2
\r
118 strneb r4, [r1], #1
\r
121 @ TileSingleHi (r1=pdest, r2=pixels8, r3=pal) r4,r7: scratch, r0=sx, r12: register with helper pattern 0xf
\r
122 .macro TileSingleHi
\r
123 tst r1, #1 @ not aligned?
\r
126 ldrneb r4, [r1], #1
\r
127 ldreqh r4, [r1], #2 @ 1ci
\r
129 bic r4, r4, r7, lsr #1
\r
131 strneb r4, [r1, #-3]
\r
132 streqh r4, [r1, #-4]
\r
134 bic r12, r12, r7, lsr #1
\r
136 strh r12, [r1, #-4]
\r
138 bic r4, r4, r7, lsr #1
\r
142 bic r12, r12, r7, lsr #1
\r
144 strh r12, [r1, #-2]
\r
145 bicne r4, r4, r7, lsr #1
\r
147 strneb r4, [r1], #1
\r
151 .macro TileDoShGenPixel shift ofs
\r
153 ands r4, r12, r2, lsr #\shift
\r
159 ldrgeb r7, [r1,#\ofs]
\r
160 orrlt r7, r3, r4 @ normal
\r
162 bicge r7, r7, #0xc0
\r
163 orrge r7, r7, r4, lsl #6
\r
164 strb r7, [r1,#\ofs]
\r
168 @ TileFlipSh (r1=pdest, r2=pixels8, r3=pal) r4,r7: scratch, r0=sx, r12: register with helper pattern 0xf
\r
170 TileDoShGenPixel 16, 0 @ #0x000f0000
\r
171 TileDoShGenPixel 20, 1 @ #0x00f00000
\r
172 TileDoShGenPixel 24, 2 @ #0x0f000000
\r
173 TileDoShGenPixel 28, 3 @ #0xf0000000
\r
174 TileDoShGenPixel 0, 4 @ #0x0000000f
\r
175 TileDoShGenPixel 4, 5 @ #0x000000f0
\r
176 TileDoShGenPixel 8, 6 @ #0x00000f00
\r
177 TileDoShGenPixel 12, 7 @ #0x0000f000
\r
180 @ TileNormSh (r1=pdest, r2=pixels8, r3=pal) r4,r7: scratch, r0=sx, r12: register with helper pattern 0xf
\r
182 TileDoShGenPixel 12, 0 @ #0x0000f000
\r
183 TileDoShGenPixel 8, 1 @ #0x00000f00
\r
184 TileDoShGenPixel 4, 2 @ #0x000000f0
\r
185 TileDoShGenPixel 0, 3 @ #0x0000000f
\r
186 TileDoShGenPixel 28, 4 @ #0xf0000000
\r
187 TileDoShGenPixel 24, 5 @ #0x0f000000
\r
188 TileDoShGenPixel 20, 6 @ #0x00f00000
\r
189 TileDoShGenPixel 16, 7 @ #0x000f0000
\r
192 .macro TileDoShGenPixel_markop shift ofs
\r
194 ands r4, r12, r2, lsr #\shift
\r
200 ldrgeb r4, [r1,#\ofs]
\r
202 orrge r4, r4, #0x80
\r
203 strb r4, [r1,#\ofs]
\r
207 .macro TileFlipSh_markop
\r
208 TileDoShGenPixel_markop 16, 0 @ #0x000f0000
\r
209 TileDoShGenPixel_markop 20, 1 @ #0x00f00000
\r
210 TileDoShGenPixel_markop 24, 2 @ #0x0f000000
\r
211 TileDoShGenPixel_markop 28, 3 @ #0xf0000000
\r
212 TileDoShGenPixel_markop 0, 4 @ #0x0000000f
\r
213 TileDoShGenPixel_markop 4, 5 @ #0x000000f0
\r
214 TileDoShGenPixel_markop 8, 6 @ #0x00000f00
\r
215 TileDoShGenPixel_markop 12, 7 @ #0x0000f000
\r
218 .macro TileNormSh_markop
\r
219 TileDoShGenPixel_markop 12, 0 @ #0x0000f000
\r
220 TileDoShGenPixel_markop 8, 1 @ #0x00000f00
\r
221 TileDoShGenPixel_markop 4, 2 @ #0x000000f0
\r
222 TileDoShGenPixel_markop 0, 3 @ #0x0000000f
\r
223 TileDoShGenPixel_markop 28, 4 @ #0xf0000000
\r
224 TileDoShGenPixel_markop 24, 5 @ #0x0f000000
\r
225 TileDoShGenPixel_markop 20, 6 @ #0x00f00000
\r
226 TileDoShGenPixel_markop 16, 7 @ #0x000f0000
\r
229 .macro TileDoShGenPixel_onlyop_lp shift ofs
\r
231 ands r7, r12, r2, lsr #\shift
\r
235 ldrneb r4, [r1,#\ofs]
\r
240 bicne r4, r4, #0xc0
\r
241 orrne r4, r4, r7, lsl #6
\r
242 strneb r4, [r1,#\ofs]
\r
246 .macro TileFlipSh_onlyop_lp
\r
247 TileDoShGenPixel_onlyop_lp 16, 0 @ #0x000f0000
\r
248 TileDoShGenPixel_onlyop_lp 20, 1 @ #0x00f00000
\r
249 TileDoShGenPixel_onlyop_lp 24, 2 @ #0x0f000000
\r
250 TileDoShGenPixel_onlyop_lp 28, 3 @ #0xf0000000
\r
251 TileDoShGenPixel_onlyop_lp 0, 4 @ #0x0000000f
\r
252 TileDoShGenPixel_onlyop_lp 4, 5 @ #0x000000f0
\r
253 TileDoShGenPixel_onlyop_lp 8, 6 @ #0x00000f00
\r
254 TileDoShGenPixel_onlyop_lp 12, 7 @ #0x0000f000
\r
257 .macro TileNormSh_onlyop_lp
\r
258 TileDoShGenPixel_onlyop_lp 12, 0 @ #0x0000f000
\r
259 TileDoShGenPixel_onlyop_lp 8, 1 @ #0x00000f00
\r
260 TileDoShGenPixel_onlyop_lp 4, 2 @ #0x000000f0
\r
261 TileDoShGenPixel_onlyop_lp 0, 3 @ #0x0000000f
\r
262 TileDoShGenPixel_onlyop_lp 28, 4 @ #0xf0000000
\r
263 TileDoShGenPixel_onlyop_lp 24, 5 @ #0x0f000000
\r
264 TileDoShGenPixel_onlyop_lp 20, 6 @ #0x00f00000
\r
265 TileDoShGenPixel_onlyop_lp 16, 7 @ #0x000f0000
\r
269 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
273 @ int nametab; // 0x00
\r
274 @ int line; // 0x04
\r
275 @ int hscroll; // 0x08
\r
276 @ int xmask; // 0x0C
\r
277 @ int *hc; // 0x10 (pointer to cache buffer)
\r
278 @ int cells; // 0x14
\r
281 @ void DrawLayer(int lflags, int *hcache, int cellskip, int maxcells,
\r
282 @ struct PicoEState *est)
\r
287 ldr r12, [sp] @ est
\r
288 stmfd sp!, {r4-r11,lr}
\r
290 ldr r11, [r12, #OFS_EST_Pico]
\r
293 ldrb r7, [r11, #OFS_Pico_video_reg+16] @ ??vv??hh
\r
295 mov r6, r1 @ hcache
\r
296 orr r9, r3, r0, lsl #29 @ r9=force[31]|sh[30]|plane[29]
\r
297 orr r9, r9, r2, lsl #8 @ |cellskip[15:8]|maxcells[7:0] (tmp)
\r
300 orr r1, r1, #0x00ff
\r
304 biclt r1, r1, #0xfc00
\r
305 biceq r1, r1, #0xfe00
\r
308 movgt r1, #0x00ff @ r1=ymask=(height<<8)|0xff; ...; // Y Mask in pixels
\r
312 subge r10, r10, #1 @ r10=shift[width] (5,6,6,7)
\r
314 ldr r2, [r12, #OFS_EST_DrawScanline]
\r
315 ldr lr, [r12, #OFS_EST_PicoMem_vram]
\r
319 ldreqb r12, [r11, #OFS_Pico_video_reg+2]
\r
320 ldrneb r12, [r11, #OFS_Pico_video_reg+4]
\r
323 mov r5, r8, lsl r10
\r
324 sub r5, r5, #1 @ r5=xmask
\r
326 moveq r12, r12, lsl #10
\r
327 movne r12, r12, lsl #13
\r
328 and r12, r12, #(7<<13) @ r12=(ts->nametab<<1) (halfword compliant)
\r
330 ldrh r8, [r11, #OFS_Pico_video_reg+12]
\r
331 ldrb r7, [r11, #OFS_Pico_video_reg+11]
\r
333 mov r4, r8, lsr #8 @ pvid->reg[13]
\r
334 mov r4, r4, lsl #10 @ htab=pvid->reg[13]<<9; (halfwords)
\r
336 addne r4, r4, r2, lsl #2 @ htab+=DrawScanline<<1; // Offset by line
\r
338 biceq r4, r4, #0x1f @ htab&=~0xf; // Offset by tile
\r
339 add r4, r4, r0, lsl #1 @ htab+=plane
\r
340 bic r4, r4, #0x00ff0000 @ just in case
\r
341 ldrh r3, [lr, r4] @ r3=hscroll
\r
344 bne .DrawStrip_vsscroll
\r
346 @ Get vertical scroll value:
\r
347 add r7, lr, #0x012000
\r
348 add r7, r7, #0x000180 @ r7=PicoMem.vsram (PicoMem+0x22180)
\r
353 bne .DrawStrip_interlace
\r
356 movne r7, r7, lsr #16
\r
358 @ Find the line in the name table
\r
362 add r10, r10, #1 @ shift[width]++
\r
363 add r12, r12, r4, lsl r10 @ nametab+=(ts.line>>3)<<shift[width];
\r
365 @ ldmia r0, {r1,r2,r3,r5,r6,r9} @ r2=line, r3=ts->hscroll, r5=ts->xmask, r6=ts->hc, r9=ts->cells
\r
368 mov r10,r10, lsl #1 @ r10=ty=(ts->line&7)<<1;
\r
369 orr r10,r10, r9, lsl #24
\r
372 mov r8, r8, lsr #3 @ r8=tilex=(-ts->hscroll)>>3
\r
376 add r7, r1, #1 @ r7=dx=((ts->hscroll-1)&7)+1
\r
378 movs r3, r9, lsl #1 @ (force[31]|sh[30]) << 1
\r
380 orrmi r10,r10, #1<<23 @ r10=cells[31:24]|sh[23]|hi_not_empty[22]
\r
381 orrcs r10,r10, #1<<20 @ |had_output[21]|force[20]|ty[15:0]
\r
382 movmi r3, #0x40 @ default to shadowed pal on sh mode
\r
385 addne r10,r10, #0x01000000 @ we will loop cells+1 times if there is scroll
\r
387 and r9, r9, #0xff00
\r
388 add r8, r8, r9, lsr #8 @ tilex+=cellskip
\r
389 add r7, r7, r9, lsr #5 @ dx+=cellskip<<3;
\r
390 sub r10,r10,r9, lsl #16 @ cells-=cellskip
\r
392 @ cache some stuff to avoid mem access
\r
393 ldr r11,[sp, #9*4] @ est
\r
395 ldr r11,[r11, #OFS_EST_HighCol]
\r
397 mvn r9, #0 @ r9=prevcode=-1
\r
398 add r1, r11, r7 @ r1=pdest
\r
401 @ r4 & r7 are scratch in this loop
\r
404 .dsloop: @ 40-41 times
\r
405 subs r10,r10, #0x01000000
\r
410 add r7, lr, r7, lsl #1 @ Pico.vram+((tilex&ts->xmask) as halfwords)
\r
411 ldrh r7, [r7, r12] @ r7=code (int, but from unsigned, no sign extend)
\r
417 tsteq r10, #1<<20 @ force?
\r
418 bne .DrawStrip_hiprio
\r
421 beq .DrawStrip_samecode @ we know stuff about this tile already
\r
423 mov r9, r7 @ remember code
\r
424 orr r10, r10, #1<<21 @ seen non hi-prio tile
\r
426 movs r2, r9, lsl #20 @ if (code&0x1000)
\r
428 add r2, r2, r10, lsl #17
\r
429 mov r2, r2, lsr #17
\r
430 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
432 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
435 and r3, r9, #0x6000
\r
436 add r3, r7, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);
\r
438 .DrawStrip_samecode:
\r
440 beq .dsloop @ tileline blank
\r
443 beq .DrawStrip_SingleColor @ tileline singlecolor
\r
446 bne .DrawStrip_TileFlip
\r
448 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r0: helper pattern
\r
449 .DrawStrip_TileNorm:
\r
453 .DrawStrip_TileFlip:
\r
457 .DrawStrip_SingleColor:
\r
460 orr r4, r4, r4, lsl #8
\r
461 tst r1, #1 @ not aligned?
\r
462 strneb r4, [r1], #1
\r
463 streqh r4, [r1], #2
\r
467 strneb r4, [r1], #1 @ have a remaining unaligned pixel?
\r
470 .DrawStrip_hiprio_maybempt:
\r
472 beq .dsloop @ must've been empty, otherwise we wouldn't get here
\r
473 movs r2, r7, lsl #20 @ if (code&0x1000)
\r
475 add r2, r2, r10, lsl #17
\r
476 mov r2, r2, lsr #17
\r
477 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
478 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
479 mov r9, r7 @ remember code
\r
482 orr r10, r10, #1<<22
\r
485 tst r10, #0x00d00000 @ sh[23]|hi_not_empty[22]|force[20]
\r
486 beq .DrawStrip_hiprio_maybempt
\r
488 orr r7, r7, r0, lsl #16
\r
489 orr r7, r7, r10, lsl #25 @ (ty<<25)
\r
491 eorne r7, r7, #7<<26 @ if(code&0x1000) cval^=7<<26;
\r
492 str r7, [r6], #4 @ cache hi priority tile
\r
497 tst r10, #1<<21 @ seen non hi-prio tile
\r
498 ldr r1, [sp, #9*4] @ est
\r
500 ldreq r2, [r1, #OFS_EST_rendstatus]
\r
501 str r0, [r6] @ terminate the cache list
\r
502 orreq r2, r2, #PDRAW_PLANE_HI_PRIO @ had a layer with all hi-prio tiles
\r
503 streq r2, [r1, #OFS_EST_rendstatus]
\r
505 ldmfd sp!, {r4-r11,lr}
\r
508 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
510 .DrawStrip_vsscroll:
\r
512 mov r8, r8, lsr #3 @ r8=tilex=(-ts->hscroll)>>3
\r
513 bic r8, r8, #0x3fc00000
\r
514 orr r8, r8, r5, lsl #25 @ r8=(xmask[31:25]|had_output[24]|tilex[21:0])
\r
516 ldr r11, [sp, #9*4] @ est
\r
517 orr r5, r1, r10, lsl #24
\r
518 ldr r4, [r11, #OFS_EST_DrawScanline]
\r
520 orr r5, r5, r4, lsl #16 @ r5=(shift_width[31:24]|scanline[23:16]|ymask[15:0])
\r
522 add r7, r1, #1 @ r7=dx=((ts->hscroll-1)&7)+1
\r
524 mov r10,r9, lsl #16
\r
526 orrne r10,r10, #0x8000
\r
529 orr r10,r10, #0xff000000 @ will be adjusted on entering loop
\r
530 orrne r10,r10, #1<<23 @ r10=(cell[31:24]|sh[23]|hi_not_empty[22]|cells_max[21:16]|plane[15]|ty[14:0])
\r
531 movne r3, #0x40 @ default to shadowed pal on sh mode
\r
534 subne r10,r10, #0x01000000 @ have hscroll, start with negative cell
\r
536 and r9, r9, #0xff00
\r
537 add r8, r8, r9, lsr #8 @ tilex+=cellskip
\r
538 add r7, r7, r9, lsr #5 @ dx+=cellskip<<3;
\r
539 add r10,r10,r9, lsl #16 @ cell+=cellskip
\r
541 @ cache some stuff to avoid mem access
\r
542 ldr r11,[sp, #9*4] @ est
\r
544 ldr r11,[r11, #OFS_EST_HighCol]
\r
546 mvn r9, #0 @ r9=prevcode=-1
\r
547 add r1, r11, r7 @ r1=pdest
\r
549 @ r4 & r7 are scratch in this loop
\r
552 .dsloop_vs: @ 40-41 times
\r
553 add r10,r10, #0x01000000
\r
554 and r4, r10, #0x003f0000
\r
555 cmp r4, r10, asr #8
\r
556 ble .dsloop_vs_exit
\r
558 @ calc offset and read tileline code to r7, also calc ty
\r
559 add r7, lr, #0x012000
\r
560 add r7, r7, #0x000180 @ r7=PicoMem.vsram (PicoMem+0x22180)
\r
561 add r7, r7, r10,asr #23 @ vsram + ((cell&~1)<<1)
\r
563 tst r10,#0x8000 @ plane1?
\r
565 ldrh r7, [r7] @ r7=vscroll
\r
567 bic r10,r10,#0xff @ clear old ty
\r
568 and r4, r5, #0xff0000 @ scanline
\r
569 add r4, r4, r7, lsl #16 @ ... += vscroll
\r
570 and r4, r4, r5, lsl #16 @ ... &= ymask
\r
571 and r7, r4, #0x70000
\r
572 orr r10,r10,r7, lsr #15 @ new ty
\r
574 mov r4, r4, lsr #19
\r
575 mov r7, r5, lsr #24
\r
576 mov r4, r4, lsl r7 @ nametabadd
\r
578 and r7, r8, r8, lsr #25
\r
579 add r7, lr, r7, lsl #1 @ PicoMem.vram+((tilex&ts->xmask) as halfwords)
\r
580 add r7, r7, r4, lsl #1
\r
581 ldrh r7, [r7, r12] @ r7=code (int, but from unsigned, no sign extend)
\r
587 bne .DrawStrip_vs_hiprio
\r
590 beq .DrawStrip_vs_samecode @ we know stuff about this tile already
\r
592 mov r9, r7 @ remember code
\r
593 orr r8, r8, #(1<<24)@ seen non hi-prio tile
\r
595 movs r2, r9, lsl #20 @ if (code&0x1000)
\r
597 add r2, r2, r10, lsl #17
\r
598 mov r2, r2, lsr #17
\r
599 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
601 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(PicoMem.vram+addr); // Get 8 pixels
\r
604 and r3, r9, #0x6000
\r
605 add r3, r7, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);
\r
607 .DrawStrip_vs_samecode:
\r
609 beq .dsloop_vs @ tileline blank
\r
612 beq .DrawStrip_vs_SingleColor @ tileline singlecolor
\r
615 bne .DrawStrip_vs_TileFlip
\r
617 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r0: helper pattern
\r
618 .DrawStrip_vs_TileNorm:
\r
622 .DrawStrip_vs_TileFlip:
\r
626 .DrawStrip_vs_SingleColor:
\r
629 orr r4, r4, r4, lsl #8
\r
630 tst r1, #1 @ not aligned?
\r
631 strneb r4, [r1], #1
\r
632 streqh r4, [r1], #2
\r
636 strneb r4, [r1], #1 @ have a remaining unaligned pixel?
\r
639 .DrawStrip_vs_hiprio:
\r
640 tst r10, #0x00c00000
\r
641 beq .DrawStrip_vs_hiprio_maybempt
\r
643 orr r7, r7, r0, lsl #16
\r
644 orr r7, r7, r10, lsl #25 @ (ty<<25)
\r
646 eorne r7, r7, #7<<26 @ if(code&0x1000) cval^=7<<26;
\r
647 str r7, [r6], #4 @ cache hi priority tile
\r
651 .DrawStrip_vs_hiprio_maybempt:
\r
653 beq .dsloop_vs @ must've been empty, otherwise we wouldn't get here
\r
654 movs r2, r7, lsl #20 @ if (code&0x1000)
\r
656 add r2, r2, r10, lsl #17
\r
657 mov r2, r2, lsr #17
\r
658 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
659 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(PicoMem.vram+addr); // Get 8 pixels
\r
660 mov r9, r7 @ remember code
\r
662 orrne r10, r10, #1<<22
\r
663 bne .DrawStrip_vs_hiprio
\r
667 tst r8, #(1<<24) @ seen non hi-prio tile
\r
668 ldr r1, [sp, #9*4] @ est
\r
670 ldreq r2, [r1, #OFS_EST_rendstatus]
\r
671 str r0, [r6] @ terminate the cache list
\r
672 orreq r2, r2, #PDRAW_PLANE_HI_PRIO @ had a layer with all hi-prio tiles
\r
673 streq r2, [r1, #OFS_EST_rendstatus]
\r
675 ldmfd sp!, {r4-r11,lr}
\r
679 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
681 @ interlace mode 2? Sonic 2?
\r
682 .DrawStrip_interlace:
\r
684 moveq r7, r7, lsl #21
\r
685 movne r7, r7, lsl #5
\r
687 @ Find the line in the name table
\r
688 add r2, r7, r2, lsl #22 @ r2=(vscroll+(DrawScanline<<1))<<21 (11 bits);
\r
689 orr r1, r1, #0x80000000
\r
690 and r2, r2, r1, ror #10 @ &((ymask<<1)|1)<<21;
\r
691 mov r2, r2, lsr #21
\r
693 mov r12, r12, lsr #1 @ halfwords
\r
694 add r0, r12, r4, lsl r10 @ nametab+=(ts.line>>4)<<shift[width];
\r
698 stmia sp, {r0,r2,r3,r5,r6,r9}
\r
701 bl DrawStripInterlace @ struct TileStrip *ts
\r
704 ldmfd sp!, {r4-r11,lr}
\r
709 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
711 @ void BackFill(int reg7, int sh, struct PicoEState *est)
\r
716 stmfd sp!, {r4-r9,lr}
\r
718 mov r0, r0, lsl #26
\r
719 ldr lr, [r2, #OFS_EST_HighCol]
\r
720 mov r0, r0, lsr #26
\r
723 orr r0, r0, r1, lsl #6
\r
724 orr r0, r0, r0, lsl #8
\r
725 orr r0, r0, r0, lsl #16
\r
736 stmia lr!, {r0-r7} @ 10*8*4
\r
747 ldmfd sp!, {r4-r9,lr}
\r
751 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
753 @ void DrawTilesFromCache(int *hc, int sh, int rlim, struct PicoEState *est)
\r
755 .global DrawTilesFromCache
\r
757 DrawTilesFromCache:
\r
758 stmfd sp!, {r4-r9,r11,lr}
\r
760 @ cache some stuff to avoid mem access
\r
761 ldr r11,[r3, #OFS_EST_HighCol]
\r
763 ldr lr, [r3, #OFS_EST_PicoMem_vram]
\r
766 mvn r5, #0 @ r5=prevcode=-1
\r
768 orr r8, r8, r2, lsl #1
\r
769 bne .dtfc_check_rendflags
\r
773 ldr r6, [r0], #4 @ read code
\r
774 movs r1, r6, lsr #16 @ r1=dx;
\r
775 ldmeqfd sp!, {r4-r9,r11,pc} @ dx is never zero, this must be a terminator, return
\r
776 bic r4, r1, #0xfe00
\r
777 add r1, r11, r4 @ r1=pdest
\r
779 movs r7, r6, lsl #16
\r
780 bpl .dtfc_loop @ !(code & 0x8000)
\r
781 cmp r5, r7, lsr #16
\r
782 beq .dtfc_samecode @ if (code==prevcode)
\r
784 mov r5, r7, lsr #16
\r
786 mov r2, r5, lsl #21
\r
787 mov r2, r2, lsr #17 @ r2=addr=(code&0x7ff)<<4;
\r
788 add r2, r2, r6, lsr #25 @ addr+=ty
\r
790 and r3, r5, #0x6000
\r
791 mov r3, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);
\r
793 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
796 rsbs r4, r4, r8, lsr #1
\r
806 beq .dtfc_SingleColor @ tileline singlecolor
\r
811 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
823 orr r4, r4, r4, lsl #8
\r
824 tst r1, #1 @ not aligned?
\r
825 strneb r4, [r1], #1
\r
826 streqh r4, [r1], #2
\r
830 strneb r4, [r1], #1 @ have a remaining unaligned pixel?
\r
835 beq .dtfc_shadow_blank
\r
838 beq .dtfc_SingleColor @ tileline singlecolor
\r
841 bne .dtfc_TileFlipShHP
\r
843 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
844 .dtfc_TileNormShHP:
\r
848 .dtfc_TileFlipShHP:
\r
852 .dtfc_shadow_blank:
\r
857 strneb r4, [r1], #1
\r
859 orr r6, r6, r6, lsl #8
\r
875 add r4, r4, #7 @ 0-6
\r
879 mov r2, r2, ror #16
\r
880 tst r5, #0x0800 @ flipped?
\r
883 mov r2, r2, ror #16
\r
893 @ check if we have detected layer covered with hi-prio tiles:
\r
894 .dtfc_check_rendflags:
\r
895 ldr r2, [r9, #OFS_EST_rendstatus]
\r
896 tst r2, #(PDRAW_PLANE_HI_PRIO|PDRAW_SHHI_DONE)
\r
898 bic r8, r8, #1 @ sh/hi mode off
\r
899 tst r2, #PDRAW_SHHI_DONE
\r
900 bne .dtfc_loop @ already processed
\r
901 orr r2, r2, #PDRAW_SHHI_DONE
\r
902 str r2, [r9, #OFS_EST_rendstatus]
\r
907 orr r6, r6, r6, lsl #8
\r
908 orr r6, r6, r6, lsl #16
\r
910 ldmia r1, {r2,r4,r5,r7}
\r
916 stmia r1!,{r2,r4,r5,r7}
\r
917 bne .dtfc_loop_shprep
\r
919 mvn r5, #0 @ r5=prevcode=-1
\r
924 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
927 @ void DrawSpritesSHi(unsigned char *sprited, struct PicoEState *est)
\r
929 .global DrawSpritesSHi
\r
937 stmfd sp!, {r1,r4-r11,lr} @ +est
\r
938 strb r12,[r0,#2] @ set end marker
\r
939 add r10,r0, #3 @ r10=HighLnSpr end
\r
940 add r10,r10,r3 @ r10=HighLnSpr end
\r
942 ldr r11,[r1, #OFS_EST_HighCol]
\r
944 ldr lr, [r1, #OFS_EST_PicoMem_vram]
\r
949 ldrb r0, [r10,#-1]!
\r
951 ldr r1, [r7, #OFS_EST_HighPreSpr]
\r
953 ldmeqfd sp!, {r1,r4-r11,pc} @ end of list
\r
955 add r0, r1, r0, lsl #3
\r
957 ldr r9, [r0, #4] @ sprite[1]
\r
958 mov r2, r9, asr #16 @ r2=sx
\r
960 mov r9, r9, lsl #16
\r
961 mov r3, r9, lsr #31 @ priority
\r
962 mov r9, r9, lsr #16
\r
963 @ orr r9, r9, r8, lsl #31 @ r9=code|sh[31] @@ sh is always on here now
\r
964 and r4, r9, #0x6000
\r
965 orr r9, r9, r4, lsl #16
\r
966 orr r9, r9, #0x90000000 @ r9=scc1 ???? ... <code> (s=shadow/hilight, cc=pal)
\r
967 cmp r12,r9, lsr #28 @ sh/hi with pal3?
\r
968 cmpne r3, #1 @ if not, is it hi prio?
\r
969 bne DrawSpriteSHi @ non-operator low sprite, already drawn
\r
971 ldr r3, [r0] @ sprite[0]
\r
972 mov r6, r3, lsr #28
\r
973 sub r6, r6, #1 @ r6=width-1 (inc later)
\r
974 mov r5, r3, lsr #24
\r
975 and r5, r5, #7 @ r5=height
\r
977 ldr r7, [r7, #OFS_EST_DrawScanline]
\r
978 mov r0, r3, lsl #16 @ r4=sy<<16 (tmp)
\r
980 sub r7, r7, r0, asr #16 @ r7=row=DrawScanline-sy
\r
983 movne r0, r5, lsl #3
\r
985 subne r7, r0, r7 @ if (code&0x1000) row=(height<<3)-1-row; // Flip Y
\r
987 add r8, r9, r7, lsr #3 @ tile+=row>>3; // Tile number increases going down
\r
989 mlane r8, r5, r6, r8 @ if (code&0x0800) { tile+=delta*(width-1);
\r
990 rsbne r5, r5, #0 @ delta=-delta; } // r5=delta now
\r
992 mov r8, r8, lsl #21
\r
993 mov r8, r8, lsr #17
\r
995 add r8, r8, r7, lsl #1 @ tile+=(row&7)<<1; // Tile address
\r
997 mov r5, r5, lsl #4 @ delta<<=4; // Delta of address
\r
998 mov r3, r4, lsr #9 @ r3=pal=((code>>9)&0x30);
\r
1000 add r6, r6, #1 @ inc now
\r
1001 adds r0, r2, #0 @ mov sx to r0 and set ZV flags
\r
1002 b .dsprShi_loop_enter
\r
1005 subs r6, r6, #1 @ width--
\r
1007 adds r0, r0, #8 @ sx+=8
\r
1008 add r8, r8, r5 @ tile+=delta
\r
1010 .dsprShi_loop_enter:
\r
1011 ble .dsprShi_loop @ sx <= 0
\r
1015 mov r8, r8, lsl #17
\r
1016 mov r8, r8, lsr #17 @ tile&=0x7fff; // Clip tile address
\r
1018 ldr r2, [lr, r8, lsl #1] @ pack=*(unsigned int *)(PicoMem.vram+addr); // Get 8 pixels
\r
1019 add r1, r11, r0 @ r1=pdest
\r
1023 cmp r12, r9, lsr #28
\r
1024 beq .dsprShi_shadow
\r
1026 cmp r2, r2, ror #4
\r
1027 beq .dsprShi_SingleColor @ tileline singlecolor
\r
1030 bne .dsprShi_TileFlip
\r
1032 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
1034 .dsprShi_TileNorm:
\r
1038 .dsprShi_TileFlip:
\r
1042 .dsprShi_SingleColor:
\r
1045 orr r4, r4, r4, lsl #8
\r
1046 tst r0, #1 @ not aligned?
\r
1047 strneb r4, [r1], #1
\r
1048 streqh r4, [r1], #2
\r
1052 strneb r4, [r1], #1
\r
1057 beq .dsprShi_shadow_lowpri
\r
1059 cmp r2, r2, ror #4
\r
1060 beq .dsprShi_singlec_sh
\r
1063 bne .dsprShi_TileFlip_sh
\r
1065 @ (r1=pdest, r2=pixels8, r3=pal) r4, r7: scratch, r12: helper pattern
\r
1066 .dsprShi_TileNorm_sh:
\r
1070 .dsprShi_TileFlip_sh:
\r
1074 .dsprShi_singlec_sh:
\r
1075 cmp r2, #0xe0000000
\r
1076 bcc .dsprShi_SingleColor @ normal singlecolor tileline (carry inverted in ARM)
\r
1077 tst r2, #0x10000000
\r
1078 bne .dsprShi_sh_sh
\r
1086 .dsprShi_shadow_lowpri:
\r
1088 bne .dsprShi_TileFlip_sh_lp
\r
1090 .dsprShi_TileNorm_sh_lp:
\r
1091 TileNormSh_onlyop_lp
\r
1094 .dsprShi_TileFlip_sh_lp:
\r
1095 TileFlipSh_onlyop_lp
\r
1100 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1102 @ void DrawAllSprites(unsigned char *sprited, int prio, int sh,
\r
1103 @ struct PicoEState *est)
\r
1105 .global DrawAllSprites
\r
1108 orr r1, r2, r1, lsl #1
\r
1110 ands r2, r2, #0x7f
\r
1113 @ time to do some real work
\r
1114 stmfd sp!, {r1,r3-r11,lr} @ +sh|prio<<1 +est
\r
1116 strb r12,[r0,#2] @ set end marker
\r
1118 add r10,r10,r2 @ r10=HighLnSpr end
\r
1120 ldr r11,[r3, #OFS_EST_HighCol]
\r
1122 ldr lr, [r3, #OFS_EST_PicoMem_vram]
\r
1124 @ + 0 : hhhhvvvv ----hhvv yyyyyyyy yyyyyyyy // v, h: horiz. size
\r
1125 @ + 4 : xxxxxxxx xxxxxxxx pccvhnnn nnnnnnnn // x: x coord + 8
\r
1128 @ draw next sprite
\r
1129 ldrb r0, [r10,#-1]!
\r
1130 ldr r8, [sp] @ sh|prio<<1
\r
1131 ldr r7, [sp, #4] @ est
\r
1132 mov r2, r0, lsr #7
\r
1134 ldmeqfd sp!, {r1,r3-r11,pc} @ end of list
\r
1135 cmp r2, r8, lsr #1
\r
1136 bne DrawSprite @ wrong priority
\r
1137 ldr r1, [r7, #OFS_EST_HighPreSpr]
\r
1139 add r0, r1, r0, lsl #3
\r
1141 ldr r3, [r0] @ sprite[0]
\r
1142 ldr r7, [r7, #OFS_EST_DrawScanline]
\r
1143 mov r6, r3, lsr #28
\r
1144 sub r6, r6, #1 @ r6=width-1 (inc later)
\r
1145 mov r5, r3, lsr #24
\r
1146 and r5, r5, #7 @ r5=height
\r
1148 mov r4, r3, lsl #16 @ r4=sy<<16 (tmp)
\r
1151 sub r7, r7, r4, asr #16 @ r7=row=DrawScanline-sy
\r
1153 mov r2, r9, asr #16 @ r2=sx
\r
1154 mov r9, r9, lsl #16
\r
1155 mov r9, r9, lsr #16
\r
1156 orr r9, r9, r8, lsl #31 @ r9=code|sh[31]
\r
1159 movne r4, r5, lsl #3
\r
1161 subne r7, r4, r7 @ if (code&0x1000) row=(height<<3)-1-row; // Flip Y
\r
1163 add r8, r9, r7, lsr #3 @ tile+=row>>3; // Tile number increases going down
\r
1165 mlane r8, r5, r6, r8 @ if (code&0x0800) { tile+=delta*(width-1);
\r
1166 rsbne r5, r5, #0 @ delta=-delta; } // r5=delta now
\r
1168 mov r8, r8, lsl #21
\r
1169 mov r8, r8, lsr #17
\r
1171 add r8, r8, r7, lsl #1 @ tile+=(row&7)<<1; // Tile address
\r
1174 @ cache some stuff to avoid mem access
\r
1175 mov r5, r5, lsl #4 @ delta<<=4; // Delta of address
\r
1176 and r4, r9, #0x6000
\r
1177 orr r9, r9, r4, lsl #16
\r
1178 orrs r9, r9, #0x10000000 @ r9=scc1 ???? ... <code> (s=shadow/hilight, cc=pal)
\r
1180 mov r3, r4, lsr #9 @ r3=pal=((code>>9)&0x30);
\r
1181 orrmi r3, r3, #0x40 @ for sh/hi
\r
1183 add r6, r6, #1 @ inc now
\r
1184 adds r0, r2, #0 @ mov sx to r0 and set ZV flags
\r
1185 b .dspr_loop_enter
\r
1188 subs r6, r6, #1 @ width--
\r
1190 adds r0, r0, #8 @ sx+=8
\r
1191 add r8, r8, r5 @ tile+=delta
\r
1194 ble .dspr_loop @ sx <= 0
\r
1198 mov r8, r8, lsl #17
\r
1199 mov r8, r8, lsr #17 @ tile&=0x7fff; // Clip tile address
\r
1201 ldr r2, [lr, r8, lsl #1] @ pack=*(unsigned int *)(PicoMem.vram+addr); // Get 8 pixels
\r
1202 add r1, r11, r0 @ r1=pdest
\r
1206 cmp r12, r9, lsr #28
\r
1209 cmp r2, r2, ror #4
\r
1210 beq .dspr_SingleColor @ tileline singlecolor
\r
1213 bne .dspr_TileFlip
\r
1215 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
1226 cmp r2, #0xe0000000
\r
1227 bcs .dspr_TileNorm_sh @ op. tileline, markop. XXX: maybe add a spec. handler?
\r
1229 .dspr_SingleColor:
\r
1232 orr r4, r4, r4, lsl #8
\r
1233 tst r0, #1 @ not aligned?
\r
1234 strneb r4, [r1], #1
\r
1235 streqh r4, [r1], #2
\r
1239 strneb r4, [r1], #1
\r
1243 cmp r2, r2, ror #4
\r
1244 beq .dspr_singlec_sh
\r
1247 bne .dspr_TileFlip_sh
\r
1249 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
1250 .dspr_TileNorm_sh:
\r
1254 .dspr_TileFlip_sh:
\r
1259 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1261 @ void DrawWindow(int tstart, int tend, int prio, int sh
\r
1262 @ struct PicoEState *est)
\r
1264 .global DrawWindow
\r
1267 ldr r12, [sp] @ est
\r
1268 stmfd sp!, {r4-r11,lr}
\r
1270 ldr r6, [r12, #OFS_EST_Pico]
\r
1271 ldr r10, [r12, #OFS_EST_DrawScanline]
\r
1272 mov r11, r12 @ est
\r
1273 ldrb r12, [r6, #OFS_Pico_video_reg+3] @ pvid->reg[3]
\r
1275 ldr r4, [r6, #OFS_Pico_video_reg+12]
\r
1276 mov r5, r10, lsr #3
\r
1278 mov r10, r10, lsl #1 @ r10=ty
\r
1280 mov r12, r12, lsl #10
\r
1282 tst r4, #1 @ 40 cell mode?
\r
1283 andne r12, r12, #0xf000 @ 0x3c<<10
\r
1284 andeq r12, r12, #0xf800
\r
1285 addne r12, r12, r5, lsl #7
\r
1286 addeq r12, r12, r5, lsl #6 @ nametab
\r
1287 add r12, r12, r0, lsl #2 @ +starttile
\r
1289 ldr lr, [r11, #OFS_EST_PicoMem_vram]
\r
1290 ldr r6, [r11, #OFS_EST_rendstatus]
\r
1292 @ fetch the first code now
\r
1293 ldrh r7, [lr, r12]
\r
1295 ands r6, r6, #PDRAW_WND_DIFF_PRIO
\r
1298 eoreq r8, r2, r7, lsr #15 @ do prio bits differ?
\r
1300 ldmeqfd sp!, {r4-r11,pc} @ yes, assume that whole window uses same priority
\r
1302 orr r6, r6, r3, lsl #8 @ shadow mode
\r
1306 @ cache some stuff to avoid mem access
\r
1307 ldr r11, [r11, #OFS_EST_HighCol]
\r
1308 mov r8, r8, lsl #1 @ cells
\r
1310 mvn r9, #0 @ r9=prevcode=-1
\r
1311 add r1, r11, r0, lsl #4 @ r1=pdest
\r
1315 @ r4,r5 are scratch in this loop
\r
1319 add r12, r12, #2 @ halfwords
\r
1320 ldrh r7, [lr, r12] @ r7=code (int, but from unsigned, no sign extend)
\r
1322 beq .dwloop_end @ done
\r
1324 eor r5, r6, r7, lsr #15
\r
1326 orrne r6, r6, #2 @ wrong pri
\r
1330 beq .dw_samecode @ we know stuff about this tile already
\r
1333 mov r9, r7 @ remember code
\r
1335 movs r2, r9, lsl #20 @ if (code&0x1000)
\r
1336 mov r2, r2, lsl #1
\r
1337 add r2, r10, r2, lsr #17 @ r2=addr=(code&0x7ff)<<4; addr+=ty
\r
1338 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
1340 and r3, r9, #0x6000
\r
1341 mov r3, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);
\r
1343 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
1350 beq .dwloop @ tileline blank
\r
1352 cmp r2, r2, ror #4
\r
1353 beq .dw_SingleColor @ tileline singlecolor
\r
1358 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r0: helper pattern
\r
1368 and r4, r0, r2 @ #0x0000000f
\r
1370 orr r4, r4, r4, lsl #8
\r
1371 orr r4, r4, r4, lsl #16
\r
1373 stmia r1!, {r4,r5}
\r
1374 b .dwloop_nor1 @ we incremeted r1 ourselves
\r
1377 tst r6, #1 @ hi pri?
\r
1378 orreq r3, r3, #0x40
\r
1379 beq .dw_shadow_done
\r
1382 orr r5, r5, r5, lsl #8
\r
1383 orr r5, r5, r5, lsl #16
\r
1392 and r2, r6, #PDRAW_WND_DIFF_PRIO
\r
1393 ldmfd sp!, {r4-r11,lr}
\r
1395 ldr r1, [r0, #OFS_EST_rendstatus]
\r
1397 str r1, [r0, #OFS_EST_rendstatus]
\r
1402 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1405 @ hilights 2 pixels in RGB444/BGR444 format
\r
1406 .macro TileDoShHi2Pixels444 reg
\r
1407 mov \reg, \reg, ror #12
\r
1408 adds \reg, \reg, #0x40000000
\r
1409 orrcs \reg, \reg, #0xf0000000
\r
1410 mov \reg, \reg, ror #28
\r
1411 adds \reg, \reg, #0x40000000
\r
1412 orrcs \reg, \reg, #0xf0000000
\r
1413 mov \reg, \reg, ror #28
\r
1414 adds \reg, \reg, #0x40000000
\r
1415 orrcs \reg, \reg, #0xf0000000
\r
1416 mov \reg, \reg, ror #24
\r
1417 adds \reg, \reg, #0x40000000
\r
1418 orrcs \reg, \reg, #0xf0000000
\r
1419 mov \reg, \reg, ror #28
\r
1420 adds \reg, \reg, #0x40000000
\r
1421 orrcs \reg, \reg, #0xf0000000
\r
1422 mov \reg, \reg, ror #28
\r
1423 adds \reg, \reg, #0x40000000
\r
1424 orrcs \reg, \reg, #0xf0000000
\r
1425 mov \reg, \reg, ror #12
\r
1429 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1432 @ Convert 0000bbb0 ggg0rrr0
\r
1433 @ to rrrrrggg gggbbbbb
\r
1435 @ r2,r3 - scratch, lr = 0x001c001c, r8 = 0x08610861
\r
1436 .macro convRGB565 reg
\r
1437 and r2, lr, \reg,lsr #7 @ b
\r
1438 and r3, lr, \reg,lsr #3 @ g
\r
1439 and \reg, lr, \reg,lsl #1 @ r
\r
1440 orr r2, r2, r3, lsl #6
\r
1441 orr \reg, r2, \reg,lsl #11
\r
1443 and r2, r8, \reg,lsr #4
\r
1444 orr \reg, \reg, r2
\r
1447 @ trashes: r2-r8,r12,lr; r8 = 0x08610861; r0,r1 are advanced
\r
1448 .macro vidConvCpyRGB565_local
\r
1449 mov r12, r2, lsr #3 @ repeats
\r
1450 mov lr, #0x001c0000
\r
1451 orr lr, lr, #0x01c @ lr == pattern 0x001c001c
\r
1454 ldmia r1!, {r4-r7}
\r
1469 .global vidConvCpyRGB565
\r
1471 vidConvCpyRGB565: @ void *to, void *from, int pixels
\r
1472 stmfd sp!, {r4-r9,lr}
\r
1474 orr r8, r8, #0x0800
\r
1475 orr r8, r8, r8, lsl #16
\r
1476 vidConvCpyRGB565_local
\r
1477 ldmfd sp!, {r4-r9,lr}
\r
1481 @ void PicoDoHighPal555(int sh, int line, struct PicoEState *est)
\r
1483 .global PicoDoHighPal555
\r
1486 stmfd sp!, {r4-r10,lr}
\r
1489 ldr r8, [r10, #OFS_EST_Pico]
\r
1491 PicoDoHighPal555_nopush:
\r
1492 orr r9, r1, r0, lsl #31 @ 0:called from FinalizeLine555, 31: s/h
\r
1494 add r0, r10, #OFS_EST_HighPal
\r
1497 strb r1, [r8, #OFS_Pico_m_dirtyPal]
\r
1499 ldr r1, [r10, #OFS_EST_PicoMem_cram]
\r
1502 orr r8, r8, #0x0800
\r
1503 orr r8, r8, r8, lsl #16
\r
1505 vidConvCpyRGB565_local
\r
1508 beq PicoDoHighPal555_end
\r
1510 add r3, r10, #OFS_EST_HighPal
\r
1512 @ shadowed pixels:
\r
1514 add r4, r3, #0x40*2
\r
1515 orr r12,r12,#0x7300
\r
1516 add r5, r3, #0xc0*2
\r
1517 orr r12,r12,r12,lsl #16
\r
1519 .fl_loopcpRGB555_sh:
\r
1520 ldmia r3!, {r1,r6}
\r
1522 and r1, r12, r1, lsr #1
\r
1523 and r6, r12, r6, lsr #1
\r
1524 stmia r4!, {r1,r6}
\r
1525 stmia r5!, {r1,r6}
\r
1526 bne .fl_loopcpRGB555_sh
\r
1528 @ hilighted pixels:
\r
1529 @ t = ((dpal[i] >> 1) & 0x738e738e) + 0x738e738e;
\r
1530 @ t |= (t >> 4) & 0x08610861;
\r
1532 sub r3, r3, #0x40*2
\r
1534 .fl_loopcpRGB555_hi:
\r
1535 ldmia r3!, {r1,r6}
\r
1536 and r1, r12, r1, lsr #1
\r
1537 and r6, r12, r6, lsr #1
\r
1540 and r5, r8, r1, lsr #4
\r
1541 and r7, r8, r6, lsr #4
\r
1544 stmia r4!, {r1,r6}
\r
1546 bne .fl_loopcpRGB555_hi
\r
1549 PicoDoHighPal555_end:
\r
1551 ldmeqfd sp!, {r4-r10,pc}
\r
1553 ldr r8, [r10, #OFS_EST_Pico]
\r
1554 b FinalizeLineRGB555_pal_done
\r
1557 @ void FinalizeLine555(int sh, int line, struct PicoEState *est)
\r
1559 .global FinalizeLine555
\r
1562 stmfd sp!, {r4-r10,lr}
\r
1564 ldr r8, [r10, #OFS_EST_Pico]
\r
1566 ldrb r2, [r8, #OFS_Pico_m_dirtyPal]
\r
1569 bne PicoDoHighPal555_nopush
\r
1571 FinalizeLineRGB555_pal_done:
\r
1572 add r3, r10, #OFS_EST_HighPal
\r
1574 ldr r12, [r10, #OFS_EST_rendstatus]
\r
1575 eors r0, r0, #1 @ sh is 0
\r
1577 tstne r12,#PDRAW_ACC_SPRITES
\r
1580 ldr r1, [r10, #OFS_EST_HighCol]
\r
1581 ldr r0, [r10, #OFS_EST_DrawLineDest]
\r
1584 ldrb r12, [r8, #OFS_Pico_video_reg+12]
\r
1585 mov lr, lr, lsl #1
\r
1588 movne r2, #320/8 @ len
\r
1589 bne .fl_no32colRGB555
\r
1590 ldr r4, [r10, #OFS_EST_PicoOpt]
\r
1594 bne .fl_32scale_RGB555
\r
1596 addeq r0, r0, #32*2
\r
1598 .fl_no32colRGB555:
\r
1600 #ifdef UNALIGNED_DRAWLINEDEST
\r
1601 @ this is basically for Gizmondo, which has unaligned odd lines in the framebuffer
\r
1610 and r4, lr, r12, lsl #1
\r
1612 and r5, lr, r12, lsr #7
\r
1614 and r6, lr, r12, lsr #15
\r
1616 orr r4, r4, r5, lsl #16
\r
1618 and r5, lr, r12, lsr #23
\r
1620 and r8, lr, r7, lsl #1
\r
1622 orr r5, r6, r5, lsl #16
\r
1624 and r6, lr, r7, lsr #7
\r
1626 and r12,lr, r7, lsr #15
\r
1627 ldrh r12,[r3, r12]
\r
1628 and r7, lr, r7, lsr #23
\r
1630 orr r8, r8, r6, lsl #16
\r
1633 orr r12,r12, r7, lsl #16
\r
1635 stmia r0!, {r4,r5,r8,r12}
\r
1636 bne .fl_loopRGB555
\r
1638 ldmfd sp!, {r4-r10,lr}
\r
1642 .fl_32scale_RGB555:
\r
1643 mov r9, #0x3900 @ f800 07e0 001f | e000 0780 001c | 3800 01e0 0007
\r
1644 orr r9, r9, #0x00e7
\r
1646 #ifdef UNALIGNED_DRAWLINEDEST
\r
1648 bne .fl_32scale_RGB555u
\r
1651 .fl_loop32scale_RGB555:
\r
1655 and r4, lr, r12,lsl #1
\r
1657 and r5, lr, r12,lsr #7
\r
1659 and r4, r4, r9, lsl #2
\r
1660 orr r4, r4, r4, lsl #14 @ r4[31:16] = 1/4 pix_s 0
\r
1661 and r5, r5, r9, lsl #2
\r
1662 sub r6, r5, r5, lsr #2 @ r6 = 3/4 pix_s 1
\r
1663 add r4, r4, r6, lsl #16 @ pix_d 0, 1
\r
1664 and r6, lr, r12,lsr #15
\r
1666 and r12,lr, r12,lsr #23
\r
1667 ldrh r12,[r3, r12]
\r
1668 and r6, r6, r9, lsl #2
\r
1670 mov r5, r5, lsr #1
\r
1671 sub r6, r6, r6, lsr #2 @ r6 = 3/4 pix_s 2
\r
1672 orr r5, r5, r6, lsl #16
\r
1674 and r6, lr, r7, lsl #1
\r
1676 and r12,r12,r9, lsl #2
\r
1677 add r5, r5, r12,lsl #14 @ pix_d 2, 3
\r
1678 and r6, r6, r9, lsl #2
\r
1679 orr r6, r12,r6, lsl #16 @ pix_d 4, 5
\r
1681 and r12,lr, r7, lsr #7
\r
1682 ldrh r12,[r3, r12]
\r
1683 and r10,lr, r7, lsr #15
\r
1684 ldrh r10,[r3, r10]
\r
1685 and r12,r12,r9, lsl #2
\r
1686 sub r8, r12,r12,lsr #2 @ r8 = 3/4 pix_s 1
\r
1687 add r8, r8, r6, lsr #18
\r
1688 and r7, lr, r7, lsr #23
\r
1690 and r10,r10,r9, lsl #2
\r
1691 orr r8, r8, r10,lsl #15
\r
1692 add r8, r8, r12,lsl #15 @ pix_d 6, 7
\r
1693 sub r10,r10,r10,lsr #2 @ r10= 3/4 pix_s 2
\r
1694 and r7, r7, r9, lsl #2
\r
1695 add r10,r10,r7, lsr #2 @ += 1/4 pix_s 3
\r
1696 orr r10,r10,r7, lsl #16 @ pix_d 8, 9
\r
1700 stmia r0!, {r4,r5,r6,r8,r10}
\r
1701 bne .fl_loop32scale_RGB555
\r
1703 ldmfd sp!, {r4-r10,lr}
\r
1706 #ifdef UNALIGNED_DRAWLINEDEST
\r
1707 @ unaligned versions of loops
\r
1708 @ warning: starts drawing 2bytes before dst
\r
1711 sub r0, r0, #2 @ initial adjustment
\r
1718 and r6, lr, r12,lsl #1
\r
1720 and r5, lr, r12,lsr #7
\r
1722 orr r4, r8, r6, lsl #16
\r
1724 and r6, lr, r12,lsr #15
\r
1726 and r8, lr, r12,lsr #23
\r
1728 orr r5, r5, r6, lsl #16
\r
1730 and r6, lr, r7, lsl #1
\r
1732 and r12,lr, r7, lsr #7
\r
1733 ldrh r12,[r3, r12]
\r
1734 orr r6, r8, r6, lsl #16
\r
1736 and r8, lr, r7, lsr #15
\r
1738 and r7, lr, r7, lsr #23
\r
1741 orr r12,r12,r8, lsl #16
\r
1744 stmia r0!, {r4,r5,r6,r12}
\r
1745 bne .fl_loopRGB555u
\r
1749 ldmfd sp!, {r4-r10,lr}
\r
1753 .fl_32scale_RGB555u:
\r
1754 sub r0, r0, #2 @ initial adjustment
\r
1757 @ r9 f800 07e0 001f | e000 0780 001c | 3800 01e0 0007
\r
1758 .fl_loop32scale_RGB555u:
\r
1762 and r6, lr, r12,lsl #1
\r
1764 and r5, lr, r12,lsr #7
\r
1766 and r6, r6, r9, lsl #2
\r
1767 orr r4, r4, r6, lsl #16 @ r4 = pix_d -1, 0
\r
1769 and r5, r5, r9, lsl #2
\r
1770 sub r8, r5, r5, lsr #2 @ r8 = 3/4 pix_s 1
\r
1771 add r6, r8, r6, lsr #2 @ r6 = (1/4 pix_s 0) + (3/4 pix_s 1)
\r
1772 orr r5, r6, r5, lsl #15
\r
1774 and r6, lr, r12,lsr #15
\r
1776 and r12,lr, r12,lsr #23
\r
1777 ldrh r12,[r3, r12]
\r
1778 and r6, r6, r9, lsl #2
\r
1779 add r5, r5, r6, lsl #15 @ r5 = pix_d 1, 2
\r
1781 and r8, lr, r7, lsl #1
\r
1783 and r10,lr, r7, lsr #7
\r
1784 ldrh r10,[r3, r10]
\r
1785 and r12,r12,r9, lsl #2
\r
1786 sub r6, r6, r6, lsr #2 @ r6 = 3/4 pix_s 2
\r
1787 add r6, r6, r12,lsr #2
\r
1788 orr r6, r6, r12,lsl #16 @ r6 = pix_d 3, 4
\r
1790 and r8, r8, r9, lsl #2
\r
1791 and r10,r10,r9, lsl #2
\r
1792 sub r12,r10,r10,lsr #2 @ r12 = 3/4 pix_s 5
\r
1793 orr r8, r8, r8, lsl #14
\r
1794 add r8, r8, r12,lsl #16 @ r8 = pix_d 5, 6
\r
1795 and r12,lr, r7, lsr #15
\r
1796 ldrh r12,[r3, r12]
\r
1797 and r7, lr, r7, lsr #23
\r
1799 and r12,r12,r9, lsl #2
\r
1801 mov r10,r10, lsr #1
\r
1802 sub r12,r12,r12,lsr #2 @ r12 = 3/4 pix_s 6
\r
1803 orr r10,r10,r12,lsl #16
\r
1804 and r7, r7, r9, lsl #2
\r
1805 add r10,r10,r7, lsl #14 @ r10 = pix_d 7, 8
\r
1809 stmia r0!, {r4,r5,r6,r8,r10}
\r
1811 bne .fl_loop32scale_RGB555u
\r
1815 ldmfd sp!, {r4-r10,lr}
\r
1818 #endif /* UNALIGNED_DRAWLINEDEST */
\r
1821 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1824 .global blockcpy @ void *dst, void *src, size_t n
\r
1827 stmfd sp!, {r4,r5}
\r
1828 mov r2, r2, lsr #4
\r
1830 ldmia r1!, {r3-r5,r12}
\r
1832 stmia r0!, {r3-r5,r12}
\r
1834 ldmfd sp!, {r4,r5}
\r
1838 .global blockcpy_or @ void *dst, void *src, size_t n, int pat
\r
1841 stmfd sp!, {r4-r6}
\r
1842 orr r3, r3, r3, lsl #8
\r
1843 orr r3, r3, r3, lsl #16
\r
1844 mov r2, r2, lsr #4
\r
1846 ldmia r1!, {r4-r6,r12}
\r
1852 stmia r0!, {r4-r6,r12}
\r
1853 bne blockcpy_loop_or
\r
1854 ldmfd sp!, {r4-r6}
\r
1857 @ vim:filetype=armasm
\r