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 plane_sh, 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_Pico_video]
\r
293 ldrb r7, [r11, #16] @ ??vv??hh
\r
295 mov r6, r1 @ hcache
\r
296 orr r9, r3, r0, lsl #30
\r
297 orr r9, r9, r2, lsl #8 @ r9=sh[31]|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_DrawScanline]
\r
315 ldr lr, [r12, #OFS_Pico_vram]
\r
319 ldreqb r12, [r11, #2]
\r
320 ldrneb r12, [r11, #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, #12]
\r
331 ldrb r7, [r11, #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=Pico.vsram (Pico+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
380 orrne r10,r10, #1<<23 @ r10=(cells<<24|sh<<23|hi_not_empty<<22|had_output<<21|ty)
\r
381 movne r3, #0x40 @ default to shadowed pal on sh mode
\r
384 addne r10,r10, #0x01000000 @ we will loop cells+1 times if there is scroll
\r
386 and r9, r9, #0xff00
\r
387 add r8, r8, r9, lsr #8 @ tilex+=cellskip
\r
388 add r7, r7, r9, lsr #5 @ dx+=cellskip<<3;
\r
389 sub r10,r10,r9, lsl #16 @ cells-=cellskip
\r
391 @ cache some stuff to avoid mem access
\r
392 ldr r11,[sp, #9*4] @ est
\r
394 ldr r11,[r11, #OFS_HighCol]
\r
396 mvn r9, #0 @ r9=prevcode=-1
\r
397 add r1, r11, r7 @ r1=pdest
\r
400 @ r4 & r7 are scratch in this loop
\r
403 .dsloop: @ 40-41 times
\r
404 subs r10,r10, #0x01000000
\r
409 add r7, lr, r7, lsl #1 @ Pico.vram+((tilex&ts->xmask) as halfwords)
\r
410 ldrh r7, [r7, r12] @ r7=code (int, but from unsigned, no sign extend)
\r
416 bne .DrawStrip_hiprio
\r
419 beq .DrawStrip_samecode @ we know stuff about this tile already
\r
421 mov r9, r7 @ remember code
\r
422 orr r10, r10, #1<<21 @ seen non hi-prio tile
\r
424 movs r2, r9, lsl #20 @ if (code&0x1000)
\r
426 add r2, r2, r10, lsl #17
\r
427 mov r2, r2, lsr #17
\r
428 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
430 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
433 and r3, r9, #0x6000
\r
434 add r3, r7, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);
\r
436 .DrawStrip_samecode:
\r
438 beq .dsloop @ tileline blank
\r
441 beq .DrawStrip_SingleColor @ tileline singlecolor
\r
444 bne .DrawStrip_TileFlip
\r
446 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r0: helper pattern
\r
447 .DrawStrip_TileNorm:
\r
451 .DrawStrip_TileFlip:
\r
455 .DrawStrip_SingleColor:
\r
458 orr r4, r4, r4, lsl #8
\r
459 tst r1, #1 @ not aligned?
\r
460 strneb r4, [r1], #1
\r
461 streqh r4, [r1], #2
\r
465 strneb r4, [r1], #1 @ have a remaining unaligned pixel?
\r
468 .DrawStrip_hiprio_maybempt:
\r
470 beq .dsloop @ must've been empty, otherwise we wouldn't get here
\r
471 movs r2, r7, lsl #20 @ if (code&0x1000)
\r
473 add r2, r2, r10, lsl #17
\r
474 mov r2, r2, lsr #17
\r
475 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
476 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
477 mov r9, r7 @ remember code
\r
480 orr r10, r10, #1<<22
\r
483 tst r10, #0x00c00000
\r
484 beq .DrawStrip_hiprio_maybempt
\r
486 orr r7, r7, r0, lsl #16
\r
487 orr r7, r7, r10, lsl #25 @ (ty<<25)
\r
489 eorne r7, r7, #7<<26 @ if(code&0x1000) cval^=7<<26;
\r
490 str r7, [r6], #4 @ cache hi priority tile
\r
495 tst r10, #1<<21 @ seen non hi-prio tile
\r
496 ldr r1, [sp, #9*4] @ est
\r
498 ldreq r2, [r1, #OFS_rendstatus]
\r
499 str r0, [r6] @ terminate the cache list
\r
500 orreq r2, r2, #PDRAW_PLANE_HI_PRIO @ had a layer with all hi-prio tiles
\r
501 streq r2, [r1, #OFS_rendstatus]
\r
503 ldmfd sp!, {r4-r11,lr}
\r
506 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
508 .DrawStrip_vsscroll:
\r
510 mov r8, r8, lsr #3 @ r8=tilex=(-ts->hscroll)>>3
\r
511 bic r8, r8, #0x3fc00000
\r
512 orr r8, r8, r5, lsl #25 @ r8=(xmask[31:25]|had_output[24]|tilex[21:0])
\r
514 ldr r11, [sp, #9*4] @ est
\r
515 orr r5, r1, r10, lsl #24
\r
516 ldr r4, [r11, #OFS_DrawScanline]
\r
518 orr r5, r5, r4, lsl #16 @ r5=(shift_width[31:24]|scanline[23:16]|ymask[15:0])
\r
520 add r7, r1, #1 @ r7=dx=((ts->hscroll-1)&7)+1
\r
522 mov r10,r9, lsl #16
\r
524 orrne r10,r10, #0x8000
\r
527 orr r10,r10, #0xff000000 @ will be adjusted on entering loop
\r
528 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
529 movne r3, #0x40 @ default to shadowed pal on sh mode
\r
532 subne r10,r10, #0x01000000 @ have hscroll, start with negative cell
\r
534 and r9, r9, #0xff00
\r
535 add r8, r8, r9, lsr #8 @ tilex+=cellskip
\r
536 add r7, r7, r9, lsr #5 @ dx+=cellskip<<3;
\r
537 add r10,r10,r9, lsl #16 @ cell+=cellskip
\r
539 @ cache some stuff to avoid mem access
\r
540 ldr r11,[sp, #9*4] @ est
\r
542 ldr r11,[r11, #OFS_HighCol]
\r
544 mvn r9, #0 @ r9=prevcode=-1
\r
545 add r1, r11, r7 @ r1=pdest
\r
547 @ r4 & r7 are scratch in this loop
\r
550 .dsloop_vs: @ 40-41 times
\r
551 add r10,r10, #0x01000000
\r
552 and r4, r10, #0x003f0000
\r
553 cmp r4, r10, asr #8
\r
554 ble .dsloop_vs_exit
\r
556 @ calc offset and read tileline code to r7, also calc ty
\r
557 add r7, lr, #0x012000
\r
558 add r7, r7, #0x000180 @ r7=Pico.vsram (Pico+0x22180)
\r
559 add r7, r7, r10,asr #23 @ vsram + ((cell&~1)<<1)
\r
561 tst r10,#0x8000 @ plane1?
\r
563 ldrh r7, [r7] @ r7=vscroll
\r
565 bic r10,r10,#0xff @ clear old ty
\r
566 and r4, r5, #0xff0000 @ scanline
\r
567 add r4, r4, r7, lsl #16 @ ... += vscroll
\r
568 and r4, r4, r5, lsl #16 @ ... &= ymask
\r
569 and r7, r4, #0x70000
\r
570 orr r10,r10,r7, lsr #15 @ new ty
\r
572 mov r4, r4, lsr #19
\r
573 mov r7, r5, lsr #24
\r
574 mov r4, r4, lsl r7 @ nametabadd
\r
576 and r7, r8, r8, lsr #25
\r
577 add r7, lr, r7, lsl #1 @ Pico.vram+((tilex&ts->xmask) as halfwords)
\r
578 add r7, r7, r4, lsl #1
\r
579 ldrh r7, [r7, r12] @ r7=code (int, but from unsigned, no sign extend)
\r
585 bne .DrawStrip_vs_hiprio
\r
588 beq .DrawStrip_vs_samecode @ we know stuff about this tile already
\r
590 mov r9, r7 @ remember code
\r
591 orr r8, r8, #(1<<24)@ seen non hi-prio tile
\r
593 movs r2, r9, lsl #20 @ if (code&0x1000)
\r
595 add r2, r2, r10, lsl #17
\r
596 mov r2, r2, lsr #17
\r
597 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
599 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
602 and r3, r9, #0x6000
\r
603 add r3, r7, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);
\r
605 .DrawStrip_vs_samecode:
\r
607 beq .dsloop_vs @ tileline blank
\r
610 beq .DrawStrip_vs_SingleColor @ tileline singlecolor
\r
613 bne .DrawStrip_vs_TileFlip
\r
615 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r0: helper pattern
\r
616 .DrawStrip_vs_TileNorm:
\r
620 .DrawStrip_vs_TileFlip:
\r
624 .DrawStrip_vs_SingleColor:
\r
627 orr r4, r4, r4, lsl #8
\r
628 tst r1, #1 @ not aligned?
\r
629 strneb r4, [r1], #1
\r
630 streqh r4, [r1], #2
\r
634 strneb r4, [r1], #1 @ have a remaining unaligned pixel?
\r
637 .DrawStrip_vs_hiprio:
\r
638 tst r10, #0x00c00000
\r
639 beq .DrawStrip_vs_hiprio_maybempt
\r
641 orr r7, r7, r0, lsl #16
\r
642 orr r7, r7, r10, lsl #25 @ (ty<<25)
\r
644 eorne r7, r7, #7<<26 @ if(code&0x1000) cval^=7<<26;
\r
645 str r7, [r6], #4 @ cache hi priority tile
\r
649 .DrawStrip_vs_hiprio_maybempt:
\r
651 beq .dsloop_vs @ must've been empty, otherwise we wouldn't get here
\r
652 movs r2, r7, lsl #20 @ if (code&0x1000)
\r
654 add r2, r2, r10, lsl #17
\r
655 mov r2, r2, lsr #17
\r
656 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
657 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
658 mov r9, r7 @ remember code
\r
660 orrne r10, r10, #1<<22
\r
661 bne .DrawStrip_vs_hiprio
\r
665 tst r8, #(1<<24) @ seen non hi-prio tile
\r
666 ldr r1, [sp, #9*4] @ est
\r
668 ldreq r2, [r1, #OFS_rendstatus]
\r
669 str r0, [r6] @ terminate the cache list
\r
670 orreq r2, r2, #PDRAW_PLANE_HI_PRIO @ had a layer with all hi-prio tiles
\r
671 streq r2, [r1, #OFS_rendstatus]
\r
673 ldmfd sp!, {r4-r11,lr}
\r
677 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
679 @ interlace mode 2? Sonic 2?
\r
680 .DrawStrip_interlace:
\r
682 moveq r7, r7, lsl #21
\r
683 movne r7, r7, lsl #5
\r
685 @ Find the line in the name table
\r
686 add r2, r7, r2, lsl #22 @ r2=(vscroll+(DrawScanline<<1))<<21 (11 bits);
\r
687 orr r1, r1, #0x80000000
\r
688 and r2, r2, r1, ror #10 @ &((ymask<<1)|1)<<21;
\r
689 mov r2, r2, lsr #21
\r
691 mov r12, r12, lsr #1 @ halfwords
\r
692 add r0, r12, r4, lsl r10 @ nametab+=(ts.line>>4)<<shift[width];
\r
696 stmia sp, {r0,r2,r3,r5,r6,r9}
\r
699 bl DrawStripInterlace @ struct TileStrip *ts
\r
702 ldmfd sp!, {r4-r11,lr}
\r
707 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
709 @ void BackFill(int reg7, int sh, struct PicoEState *est)
\r
714 stmfd sp!, {r4-r9,lr}
\r
716 mov r0, r0, lsl #26
\r
717 ldr lr, [r2, #OFS_HighCol]
\r
718 mov r0, r0, lsr #26
\r
721 orr r0, r0, r1, lsl #6
\r
722 orr r0, r0, r0, lsl #8
\r
723 orr r0, r0, r0, lsl #16
\r
734 stmia lr!, {r0-r7} @ 10*8*4
\r
745 ldmfd sp!, {r4-r9,lr}
\r
749 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
751 @ void DrawTilesFromCache(int *hc, int sh, int rlim, struct PicoEState *est)
\r
753 .global DrawTilesFromCache
\r
755 DrawTilesFromCache:
\r
756 stmfd sp!, {r4-r9,r11,lr}
\r
758 @ cache some stuff to avoid mem access
\r
759 ldr r11,[r3, #OFS_HighCol]
\r
761 ldr lr, [r3, #OFS_Pico_vram]
\r
764 mvn r5, #0 @ r5=prevcode=-1
\r
766 orr r8, r8, r2, lsl #1
\r
767 bne .dtfc_check_rendflags
\r
771 ldr r6, [r0], #4 @ read code
\r
772 movs r1, r6, lsr #16 @ r1=dx;
\r
773 ldmeqfd sp!, {r4-r9,r11,pc} @ dx is never zero, this must be a terminator, return
\r
774 bic r4, r1, #0xfe00
\r
775 add r1, r11, r4 @ r1=pdest
\r
777 mov r7, r6, lsl #16
\r
778 cmp r5, r7, lsr #16
\r
779 beq .dtfc_samecode @ if (code==prevcode)
\r
781 mov r5, r7, lsr #16
\r
783 mov r2, r5, lsl #21
\r
784 mov r2, r2, lsr #17 @ r2=addr=(code&0x7ff)<<4;
\r
785 add r2, r2, r6, lsr #25 @ addr+=ty
\r
787 and r3, r5, #0x6000
\r
788 mov r3, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);
\r
790 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
793 rsbs r4, r4, r8, lsr #1
\r
803 beq .dtfc_SingleColor @ tileline singlecolor
\r
808 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
820 orr r4, r4, r4, lsl #8
\r
821 tst r1, #1 @ not aligned?
\r
822 strneb r4, [r1], #1
\r
823 streqh r4, [r1], #2
\r
827 strneb r4, [r1], #1 @ have a remaining unaligned pixel?
\r
832 beq .dtfc_shadow_blank
\r
835 beq .dtfc_SingleColor @ tileline singlecolor
\r
838 bne .dtfc_TileFlipShHP
\r
840 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
841 .dtfc_TileNormShHP:
\r
845 .dtfc_TileFlipShHP:
\r
849 .dtfc_shadow_blank:
\r
854 strneb r4, [r1], #1
\r
856 orr r6, r6, r6, lsl #8
\r
872 add r4, r4, #7 @ 0-6
\r
876 mov r2, r2, ror #16
\r
877 tst r5, #0x0800 @ flipped?
\r
880 mov r2, r2, ror #16
\r
890 @ check if we have detected layer covered with hi-prio tiles:
\r
891 .dtfc_check_rendflags:
\r
892 ldr r2, [r9, #OFS_rendstatus]
\r
893 tst r2, #(PDRAW_PLANE_HI_PRIO|PDRAW_SHHI_DONE)
\r
895 bic r8, r8, #1 @ sh/hi mode off
\r
896 tst r2, #PDRAW_SHHI_DONE
\r
897 bne .dtfc_loop @ already processed
\r
898 orr r2, r2, #PDRAW_SHHI_DONE
\r
899 str r2, [r9, #OFS_rendstatus]
\r
904 orr r6, r6, r6, lsl #8
\r
905 orr r6, r6, r6, lsl #16
\r
907 ldmia r1, {r2,r4,r5,r7}
\r
913 stmia r1!,{r2,r4,r5,r7}
\r
914 bne .dtfc_loop_shprep
\r
916 mvn r5, #0 @ r5=prevcode=-1
\r
921 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
924 @ void DrawSpritesSHi(unsigned char *sprited, struct PicoEState *est)
\r
926 .global DrawSpritesSHi
\r
934 stmfd sp!, {r1,r4-r11,lr} @ +est
\r
935 strb r12,[r0,#2] @ set end marker
\r
936 add r10,r0, #3 @ r10=HighLnSpr end
\r
937 add r10,r10,r3 @ r10=HighLnSpr end
\r
939 ldr r11,[r1, #OFS_HighCol]
\r
941 ldr lr, [r1, #OFS_Pico_vram]
\r
946 ldrb r0, [r10,#-1]!
\r
948 ldr r1, [r7, #OFS_HighPreSpr]
\r
950 ldmeqfd sp!, {r1,r4-r11,pc} @ end of list
\r
952 add r0, r1, r0, lsl #3
\r
954 ldr r9, [r0, #4] @ sprite[1]
\r
955 mov r2, r9, asr #16 @ r2=sx
\r
957 mov r9, r9, lsl #16
\r
958 mov r3, r9, lsr #31 @ priority
\r
959 mov r9, r9, lsr #16
\r
960 @ orr r9, r9, r8, lsl #31 @ r9=code|sh[31] @@ sh is always on here now
\r
961 and r4, r9, #0x6000
\r
962 orr r9, r9, r4, lsl #16
\r
963 orr r9, r9, #0x90000000 @ r9=scc1 ???? ... <code> (s=shadow/hilight, cc=pal)
\r
964 cmp r12,r9, lsr #28 @ sh/hi with pal3?
\r
965 cmpne r3, #1 @ if not, is it hi prio?
\r
966 bne DrawSpriteSHi @ non-operator low sprite, already drawn
\r
968 ldr r3, [r0] @ sprite[0]
\r
969 mov r6, r3, lsr #28
\r
970 sub r6, r6, #1 @ r6=width-1 (inc later)
\r
971 mov r5, r3, lsr #24
\r
972 and r5, r5, #7 @ r5=height
\r
974 ldr r7, [r7, #OFS_DrawScanline]
\r
975 mov r0, r3, lsl #16 @ r4=sy<<16 (tmp)
\r
977 sub r7, r7, r0, asr #16 @ r7=row=DrawScanline-sy
\r
980 movne r0, r5, lsl #3
\r
982 subne r7, r0, r7 @ if (code&0x1000) row=(height<<3)-1-row; // Flip Y
\r
984 add r8, r9, r7, lsr #3 @ tile+=row>>3; // Tile number increases going down
\r
986 mlane r8, r5, r6, r8 @ if (code&0x0800) { tile+=delta*(width-1);
\r
987 rsbne r5, r5, #0 @ delta=-delta; } // r5=delta now
\r
989 mov r8, r8, lsl #21
\r
990 mov r8, r8, lsr #17
\r
992 add r8, r8, r7, lsl #1 @ tile+=(row&7)<<1; // Tile address
\r
994 mov r5, r5, lsl #4 @ delta<<=4; // Delta of address
\r
995 mov r3, r4, lsr #9 @ r3=pal=((code>>9)&0x30);
\r
997 add r6, r6, #1 @ inc now
\r
998 adds r0, r2, #0 @ mov sx to r0 and set ZV flags
\r
999 b .dsprShi_loop_enter
\r
1002 subs r6, r6, #1 @ width--
\r
1004 adds r0, r0, #8 @ sx+=8
\r
1005 add r8, r8, r5 @ tile+=delta
\r
1007 .dsprShi_loop_enter:
\r
1008 ble .dsprShi_loop @ sx <= 0
\r
1012 mov r8, r8, lsl #17
\r
1013 mov r8, r8, lsr #17 @ tile&=0x7fff; // Clip tile address
\r
1015 ldr r2, [lr, r8, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
1016 add r1, r11, r0 @ r1=pdest
\r
1020 cmp r12, r9, lsr #28
\r
1021 beq .dsprShi_shadow
\r
1023 cmp r2, r2, ror #4
\r
1024 beq .dsprShi_SingleColor @ tileline singlecolor
\r
1027 bne .dsprShi_TileFlip
\r
1029 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
1031 .dsprShi_TileNorm:
\r
1035 .dsprShi_TileFlip:
\r
1039 .dsprShi_SingleColor:
\r
1042 orr r4, r4, r4, lsl #8
\r
1043 tst r0, #1 @ not aligned?
\r
1044 strneb r4, [r1], #1
\r
1045 streqh r4, [r1], #2
\r
1049 strneb r4, [r1], #1
\r
1054 beq .dsprShi_shadow_lowpri
\r
1056 cmp r2, r2, ror #4
\r
1057 beq .dsprShi_singlec_sh
\r
1060 bne .dsprShi_TileFlip_sh
\r
1062 @ (r1=pdest, r2=pixels8, r3=pal) r4, r7: scratch, r12: helper pattern
\r
1063 .dsprShi_TileNorm_sh:
\r
1067 .dsprShi_TileFlip_sh:
\r
1071 .dsprShi_singlec_sh:
\r
1072 cmp r2, #0xe0000000
\r
1073 bcc .dsprShi_SingleColor @ normal singlecolor tileline (carry inverted in ARM)
\r
1074 tst r2, #0x10000000
\r
1075 bne .dsprShi_sh_sh
\r
1083 .dsprShi_shadow_lowpri:
\r
1085 bne .dsprShi_TileFlip_sh_lp
\r
1087 .dsprShi_TileNorm_sh_lp:
\r
1088 TileNormSh_onlyop_lp
\r
1091 .dsprShi_TileFlip_sh_lp:
\r
1092 TileFlipSh_onlyop_lp
\r
1097 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1099 @ void DrawAllSprites(unsigned char *sprited, int prio, int sh,
\r
1100 @ struct PicoEState *est)
\r
1102 .global DrawAllSprites
\r
1105 orr r1, r2, r1, lsl #1
\r
1107 ands r2, r2, #0x7f
\r
1110 @ time to do some real work
\r
1111 stmfd sp!, {r1,r3-r11,lr} @ +sh|prio<<1 +est
\r
1113 strb r12,[r0,#2] @ set end marker
\r
1115 add r10,r10,r2 @ r10=HighLnSpr end
\r
1117 ldr r11,[r3, #OFS_HighCol]
\r
1119 ldr lr, [r3, #OFS_Pico_vram]
\r
1121 @ + 0 : hhhhvvvv ----hhvv yyyyyyyy yyyyyyyy // v, h: horiz. size
\r
1122 @ + 4 : xxxxxxxx xxxxxxxx pccvhnnn nnnnnnnn // x: x coord + 8
\r
1125 @ draw next sprite
\r
1126 ldrb r0, [r10,#-1]!
\r
1127 ldr r8, [sp] @ sh|prio<<1
\r
1128 ldr r7, [sp, #4] @ est
\r
1129 mov r2, r0, lsr #7
\r
1131 ldmeqfd sp!, {r1,r3-r11,pc} @ end of list
\r
1132 cmp r2, r8, lsr #1
\r
1133 bne DrawSprite @ wrong priority
\r
1134 ldr r1, [r7, #OFS_HighPreSpr]
\r
1136 add r0, r1, r0, lsl #3
\r
1138 ldr r3, [r0] @ sprite[0]
\r
1139 ldr r7, [r7, #OFS_DrawScanline]
\r
1140 mov r6, r3, lsr #28
\r
1141 sub r6, r6, #1 @ r6=width-1 (inc later)
\r
1142 mov r5, r3, lsr #24
\r
1143 and r5, r5, #7 @ r5=height
\r
1145 mov r4, r3, lsl #16 @ r4=sy<<16 (tmp)
\r
1148 sub r7, r7, r4, asr #16 @ r7=row=DrawScanline-sy
\r
1150 mov r2, r9, asr #16 @ r2=sx
\r
1151 mov r9, r9, lsl #16
\r
1152 mov r9, r9, lsr #16
\r
1153 orr r9, r9, r8, lsl #31 @ r9=code|sh[31]
\r
1156 movne r4, r5, lsl #3
\r
1158 subne r7, r4, r7 @ if (code&0x1000) row=(height<<3)-1-row; // Flip Y
\r
1160 add r8, r9, r7, lsr #3 @ tile+=row>>3; // Tile number increases going down
\r
1162 mlane r8, r5, r6, r8 @ if (code&0x0800) { tile+=delta*(width-1);
\r
1163 rsbne r5, r5, #0 @ delta=-delta; } // r5=delta now
\r
1165 mov r8, r8, lsl #21
\r
1166 mov r8, r8, lsr #17
\r
1168 add r8, r8, r7, lsl #1 @ tile+=(row&7)<<1; // Tile address
\r
1171 @ cache some stuff to avoid mem access
\r
1172 mov r5, r5, lsl #4 @ delta<<=4; // Delta of address
\r
1173 and r4, r9, #0x6000
\r
1174 orr r9, r9, r4, lsl #16
\r
1175 orrs r9, r9, #0x10000000 @ r9=scc1 ???? ... <code> (s=shadow/hilight, cc=pal)
\r
1177 mov r3, r4, lsr #9 @ r3=pal=((code>>9)&0x30);
\r
1178 orrmi r3, r3, #0x40 @ for sh/hi
\r
1180 add r6, r6, #1 @ inc now
\r
1181 adds r0, r2, #0 @ mov sx to r0 and set ZV flags
\r
1182 b .dspr_loop_enter
\r
1185 subs r6, r6, #1 @ width--
\r
1187 adds r0, r0, #8 @ sx+=8
\r
1188 add r8, r8, r5 @ tile+=delta
\r
1191 ble .dspr_loop @ sx <= 0
\r
1195 mov r8, r8, lsl #17
\r
1196 mov r8, r8, lsr #17 @ tile&=0x7fff; // Clip tile address
\r
1198 ldr r2, [lr, r8, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
1199 add r1, r11, r0 @ r1=pdest
\r
1203 cmp r12, r9, lsr #28
\r
1206 cmp r2, r2, ror #4
\r
1207 beq .dspr_SingleColor @ tileline singlecolor
\r
1210 bne .dspr_TileFlip
\r
1212 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
1223 cmp r2, #0xe0000000
\r
1224 bcs .dspr_TileNorm_sh @ op. tileline, markop. XXX: maybe add a spec. handler?
\r
1226 .dspr_SingleColor:
\r
1229 orr r4, r4, r4, lsl #8
\r
1230 tst r0, #1 @ not aligned?
\r
1231 strneb r4, [r1], #1
\r
1232 streqh r4, [r1], #2
\r
1236 strneb r4, [r1], #1
\r
1240 cmp r2, r2, ror #4
\r
1241 beq .dspr_singlec_sh
\r
1244 bne .dspr_TileFlip_sh
\r
1246 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
1247 .dspr_TileNorm_sh:
\r
1251 .dspr_TileFlip_sh:
\r
1256 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1258 @ void DrawWindow(int tstart, int tend, int prio, int sh
\r
1259 @ struct PicoEState *est)
\r
1261 .global DrawWindow
\r
1264 ldr r12, [sp] @ est
\r
1265 stmfd sp!, {r4-r11,lr}
\r
1267 ldr r6, [r12, #OFS_Pico_video]
\r
1268 ldr r10, [r12, #OFS_DrawScanline]
\r
1269 mov r11, r12 @ est
\r
1270 ldrb r12, [r6, #3] @ pvid->reg[3]
\r
1273 mov r5, r10, lsr #3
\r
1275 mov r10, r10, lsl #1 @ r10=ty
\r
1277 mov r12, r12, lsl #10
\r
1279 tst r4, #1 @ 40 cell mode?
\r
1280 andne r12, r12, #0xf000 @ 0x3c<<10
\r
1281 andeq r12, r12, #0xf800
\r
1282 addne r12, r12, r5, lsl #7
\r
1283 addeq r12, r12, r5, lsl #6 @ nametab
\r
1284 add r12, r12, r0, lsl #2 @ +starttile
\r
1286 ldr lr, [r11, #OFS_Pico_vram]
\r
1287 ldr r6, [r11, #OFS_rendstatus]
\r
1289 @ fetch the first code now
\r
1290 ldrh r7, [lr, r12]
\r
1292 ands r6, r6, #PDRAW_WND_DIFF_PRIO
\r
1295 eoreq r8, r2, r7, lsr #15 @ do prio bits differ?
\r
1297 ldmeqfd sp!, {r4-r11,pc} @ yes, assume that whole window uses same priority
\r
1299 orr r6, r6, r3, lsl #8 @ shadow mode
\r
1303 @ cache some stuff to avoid mem access
\r
1304 ldr r11, [r11, #OFS_HighCol]
\r
1305 mov r8, r8, lsl #1 @ cells
\r
1307 mvn r9, #0 @ r9=prevcode=-1
\r
1308 add r1, r11, r0, lsl #4 @ r1=pdest
\r
1312 @ r4,r5 are scratch in this loop
\r
1316 add r12, r12, #2 @ halfwords
\r
1317 ldrh r7, [lr, r12] @ r7=code (int, but from unsigned, no sign extend)
\r
1319 beq .dwloop_end @ done
\r
1321 eor r5, r6, r7, lsr #15
\r
1323 orrne r6, r6, #2 @ wrong pri
\r
1327 beq .dw_samecode @ we know stuff about this tile already
\r
1330 mov r9, r7 @ remember code
\r
1332 movs r2, r9, lsl #20 @ if (code&0x1000)
\r
1333 mov r2, r2, lsl #1
\r
1334 add r2, r10, r2, lsr #17 @ r2=addr=(code&0x7ff)<<4; addr+=ty
\r
1335 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
1337 and r3, r9, #0x6000
\r
1338 mov r3, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);
\r
1340 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
1347 beq .dwloop @ tileline blank
\r
1349 cmp r2, r2, ror #4
\r
1350 beq .dw_SingleColor @ tileline singlecolor
\r
1355 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r0: helper pattern
\r
1365 and r4, r0, r2 @ #0x0000000f
\r
1367 orr r4, r4, r4, lsl #8
\r
1368 orr r4, r4, r4, lsl #16
\r
1370 stmia r1!, {r4,r5}
\r
1371 b .dwloop_nor1 @ we incremeted r1 ourselves
\r
1374 tst r6, #1 @ hi pri?
\r
1375 orreq r3, r3, #0x40
\r
1376 beq .dw_shadow_done
\r
1379 orr r5, r5, r5, lsl #8
\r
1380 orr r5, r5, r5, lsl #16
\r
1389 and r2, r6, #PDRAW_WND_DIFF_PRIO
\r
1390 ldmfd sp!, {r4-r11,lr}
\r
1392 ldr r1, [r0, #OFS_rendstatus]
\r
1394 str r1, [r0, #OFS_rendstatus]
\r
1399 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1402 @ hilights 2 pixels in RGB444/BGR444 format
\r
1403 .macro TileDoShHi2Pixels444 reg
\r
1404 mov \reg, \reg, ror #12
\r
1405 adds \reg, \reg, #0x40000000
\r
1406 orrcs \reg, \reg, #0xf0000000
\r
1407 mov \reg, \reg, ror #28
\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 #24
\r
1414 adds \reg, \reg, #0x40000000
\r
1415 orrcs \reg, \reg, #0xf0000000
\r
1416 mov \reg, \reg, ror #28
\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 #12
\r
1426 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1429 @ Convert 0000bbb0 ggg0rrr0
\r
1430 @ to rrrrrggg gggbbbbb
\r
1432 @ r2,r3 - scratch, lr = 0x001c001c, r8 = 0x08610861
\r
1433 .macro convRGB565 reg
\r
1434 and r2, lr, \reg,lsr #7 @ b
\r
1435 and r3, lr, \reg,lsr #3 @ g
\r
1436 and \reg, lr, \reg,lsl #1 @ r
\r
1437 orr r2, r2, r3, lsl #6
\r
1438 orr \reg, r2, \reg,lsl #11
\r
1440 and r2, r8, \reg,lsr #4
\r
1441 orr \reg, \reg, r2
\r
1444 @ trashes: r2-r8,r12,lr; r8 = 0x08610861; r0,r1 are advanced
\r
1445 .macro vidConvCpyRGB565_local
\r
1446 mov r12, r2, lsr #3 @ repeats
\r
1447 mov lr, #0x001c0000
\r
1448 orr lr, lr, #0x01c @ lr == pattern 0x001c001c
\r
1451 ldmia r1!, {r4-r7}
\r
1466 .global vidConvCpyRGB565
\r
1468 vidConvCpyRGB565: @ void *to, void *from, int pixels
\r
1469 stmfd sp!, {r4-r9,lr}
\r
1471 orr r8, r8, #0x0800
\r
1472 orr r8, r8, r8, lsl #16
\r
1473 vidConvCpyRGB565_local
\r
1474 ldmfd sp!, {r4-r9,lr}
\r
1478 @ void PicoDoHighPal555(int sh, int line, struct PicoEState *est)
\r
1480 .global PicoDoHighPal555
\r
1483 stmfd sp!, {r4-r10,lr}
\r
1486 ldr r8, [r10, #OFS_Pico_video]
\r
1488 PicoDoHighPal555_nopush:
\r
1489 orr r9, r1, r0, lsl #31 @ 0:called from FinalizeLine555, 31: s/h
\r
1491 add r0, r10, #OFS_HighPal
\r
1494 strb r1, [r8, #-0x1a] @ 0x2220e ~ dirtyPal
\r
1496 sub r1, r8, #0x128 @ r1=Pico.cram
\r
1499 orr r8, r8, #0x0800
\r
1500 orr r8, r8, r8, lsl #16
\r
1502 vidConvCpyRGB565_local
\r
1505 beq PicoDoHighPal555_end
\r
1507 add r3, r10, #OFS_HighPal
\r
1509 @ shadowed pixels:
\r
1511 add r4, r3, #0x40*2
\r
1512 orr r12,r12,#0x7300
\r
1513 add r5, r3, #0xc0*2
\r
1514 orr r12,r12,r12,lsl #16
\r
1516 .fl_loopcpRGB555_sh:
\r
1517 ldmia r3!, {r1,r6}
\r
1519 and r1, r12, r1, lsr #1
\r
1520 and r6, r12, r6, lsr #1
\r
1521 stmia r4!, {r1,r6}
\r
1522 stmia r5!, {r1,r6}
\r
1523 bne .fl_loopcpRGB555_sh
\r
1525 @ hilighted pixels:
\r
1526 @ t = ((dpal[i] >> 1) & 0x738e738e) + 0x738e738e;
\r
1527 @ t |= (t >> 4) & 0x08610861;
\r
1529 sub r3, r3, #0x40*2
\r
1531 .fl_loopcpRGB555_hi:
\r
1532 ldmia r3!, {r1,r6}
\r
1533 and r1, r12, r1, lsr #1
\r
1534 and r6, r12, r6, lsr #1
\r
1537 and r5, r8, r1, lsr #4
\r
1538 and r7, r8, r6, lsr #4
\r
1541 stmia r4!, {r1,r6}
\r
1543 bne .fl_loopcpRGB555_hi
\r
1546 PicoDoHighPal555_end:
\r
1548 ldmeqfd sp!, {r4-r10,pc}
\r
1550 ldr r8, [r10, #OFS_Pico_video]
\r
1551 b FinalizeLineRGB555_pal_done
\r
1554 @ void FinalizeLine555(int sh, int line, struct PicoEState *est)
\r
1556 .global FinalizeLine555
\r
1559 stmfd sp!, {r4-r10,lr}
\r
1561 ldr r8, [r10, #OFS_Pico_video]
\r
1563 ldrb r2, [r8, #-0x1a] @ 0x2220e ~ dirtyPal
\r
1566 bne PicoDoHighPal555_nopush
\r
1568 FinalizeLineRGB555_pal_done:
\r
1569 add r3, r10, #OFS_HighPal
\r
1571 ldr r12, [r10, #OFS_rendstatus]
\r
1572 eors r0, r0, #1 @ sh is 0
\r
1574 tstne r12,#PDRAW_ACC_SPRITES
\r
1577 ldr r1, [r10, #OFS_HighCol]
\r
1578 ldr r0, [r10, #OFS_DrawLineDest]
\r
1581 ldrb r12, [r8, #12]
\r
1582 mov lr, lr, lsl #1
\r
1585 movne r2, #320/8 @ len
\r
1586 bne .fl_no32colRGB555
\r
1587 ldr r4, [r10, #OFS_PicoOpt]
\r
1591 bne .fl_32scale_RGB555
\r
1593 addeq r0, r0, #32*2
\r
1595 .fl_no32colRGB555:
\r
1597 #ifdef UNALIGNED_DRAWLINEDEST
\r
1598 @ this is basically for Gizmondo, which has unaligned odd lines in the framebuffer
\r
1607 and r4, lr, r12, lsl #1
\r
1609 and r5, lr, r12, lsr #7
\r
1611 and r6, lr, r12, lsr #15
\r
1613 orr r4, r4, r5, lsl #16
\r
1615 and r5, lr, r12, lsr #23
\r
1617 and r8, lr, r7, lsl #1
\r
1619 orr r5, r6, r5, lsl #16
\r
1621 and r6, lr, r7, lsr #7
\r
1623 and r12,lr, r7, lsr #15
\r
1624 ldrh r12,[r3, r12]
\r
1625 and r7, lr, r7, lsr #23
\r
1627 orr r8, r8, r6, lsl #16
\r
1630 orr r12,r12, r7, lsl #16
\r
1632 stmia r0!, {r4,r5,r8,r12}
\r
1633 bne .fl_loopRGB555
\r
1635 ldmfd sp!, {r4-r10,lr}
\r
1639 .fl_32scale_RGB555:
\r
1640 mov r9, #0x3900 @ f800 07e0 001f | e000 0780 001c | 3800 01e0 0007
\r
1641 orr r9, r9, #0x00e7
\r
1643 #ifdef UNALIGNED_DRAWLINEDEST
\r
1645 bne .fl_32scale_RGB555u
\r
1648 .fl_loop32scale_RGB555:
\r
1652 and r4, lr, r12,lsl #1
\r
1654 and r5, lr, r12,lsr #7
\r
1656 and r4, r4, r9, lsl #2
\r
1657 orr r4, r4, r4, lsl #14 @ r4[31:16] = 1/4 pix_s 0
\r
1658 and r5, r5, r9, lsl #2
\r
1659 sub r6, r5, r5, lsr #2 @ r6 = 3/4 pix_s 1
\r
1660 add r4, r4, r6, lsl #16 @ pix_d 0, 1
\r
1661 and r6, lr, r12,lsr #15
\r
1663 and r12,lr, r12,lsr #23
\r
1664 ldrh r12,[r3, r12]
\r
1665 and r6, r6, r9, lsl #2
\r
1667 mov r5, r5, lsr #1
\r
1668 sub r6, r6, r6, lsr #2 @ r6 = 3/4 pix_s 2
\r
1669 orr r5, r5, r6, lsl #16
\r
1671 and r6, lr, r7, lsl #1
\r
1673 and r12,r12,r9, lsl #2
\r
1674 add r5, r5, r12,lsl #14 @ pix_d 2, 3
\r
1675 and r6, r6, r9, lsl #2
\r
1676 orr r6, r12,r6, lsl #16 @ pix_d 4, 5
\r
1678 and r12,lr, r7, lsr #7
\r
1679 ldrh r12,[r3, r12]
\r
1680 and r10,lr, r7, lsr #15
\r
1681 ldrh r10,[r3, r10]
\r
1682 and r12,r12,r9, lsl #2
\r
1683 sub r8, r12,r12,lsr #2 @ r8 = 3/4 pix_s 1
\r
1684 add r8, r8, r6, lsr #18
\r
1685 and r7, lr, r7, lsr #23
\r
1687 and r10,r10,r9, lsl #2
\r
1688 orr r8, r8, r10,lsl #15
\r
1689 add r8, r8, r12,lsl #15 @ pix_d 6, 7
\r
1690 sub r10,r10,r10,lsr #2 @ r10= 3/4 pix_s 2
\r
1691 and r7, r7, r9, lsl #2
\r
1692 add r10,r10,r7, lsr #2 @ += 1/4 pix_s 3
\r
1693 orr r10,r10,r7, lsl #16 @ pix_d 8, 9
\r
1697 stmia r0!, {r4,r5,r6,r8,r10}
\r
1698 bne .fl_loop32scale_RGB555
\r
1700 ldmfd sp!, {r4-r10,lr}
\r
1703 #ifdef UNALIGNED_DRAWLINEDEST
\r
1704 @ unaligned versions of loops
\r
1705 @ warning: starts drawing 2bytes before dst
\r
1708 sub r0, r0, #2 @ initial adjustment
\r
1715 and r6, lr, r12,lsl #1
\r
1717 and r5, lr, r12,lsr #7
\r
1719 orr r4, r8, r6, lsl #16
\r
1721 and r6, lr, r12,lsr #15
\r
1723 and r8, lr, r12,lsr #23
\r
1725 orr r5, r5, r6, lsl #16
\r
1727 and r6, lr, r7, lsl #1
\r
1729 and r12,lr, r7, lsr #7
\r
1730 ldrh r12,[r3, r12]
\r
1731 orr r6, r8, r6, lsl #16
\r
1733 and r8, lr, r7, lsr #15
\r
1735 and r7, lr, r7, lsr #23
\r
1738 orr r12,r12,r8, lsl #16
\r
1741 stmia r0!, {r4,r5,r6,r12}
\r
1742 bne .fl_loopRGB555u
\r
1746 ldmfd sp!, {r4-r10,lr}
\r
1750 .fl_32scale_RGB555u:
\r
1751 sub r0, r0, #2 @ initial adjustment
\r
1754 @ r9 f800 07e0 001f | e000 0780 001c | 3800 01e0 0007
\r
1755 .fl_loop32scale_RGB555u:
\r
1759 and r6, lr, r12,lsl #1
\r
1761 and r5, lr, r12,lsr #7
\r
1763 and r6, r6, r9, lsl #2
\r
1764 orr r4, r4, r6, lsl #16 @ r4 = pix_d -1, 0
\r
1766 and r5, r5, r9, lsl #2
\r
1767 sub r8, r5, r5, lsr #2 @ r8 = 3/4 pix_s 1
\r
1768 add r6, r8, r6, lsr #2 @ r6 = (1/4 pix_s 0) + (3/4 pix_s 1)
\r
1769 orr r5, r6, r5, lsl #15
\r
1771 and r6, lr, r12,lsr #15
\r
1773 and r12,lr, r12,lsr #23
\r
1774 ldrh r12,[r3, r12]
\r
1775 and r6, r6, r9, lsl #2
\r
1776 add r5, r5, r6, lsl #15 @ r5 = pix_d 1, 2
\r
1778 and r8, lr, r7, lsl #1
\r
1780 and r10,lr, r7, lsr #7
\r
1781 ldrh r10,[r3, r10]
\r
1782 and r12,r12,r9, lsl #2
\r
1783 sub r6, r6, r6, lsr #2 @ r6 = 3/4 pix_s 2
\r
1784 add r6, r6, r12,lsr #2
\r
1785 orr r6, r6, r12,lsl #16 @ r6 = pix_d 3, 4
\r
1787 and r8, r8, r9, lsl #2
\r
1788 and r10,r10,r9, lsl #2
\r
1789 sub r12,r10,r10,lsr #2 @ r12 = 3/4 pix_s 5
\r
1790 orr r8, r8, r8, lsl #14
\r
1791 add r8, r8, r12,lsl #16 @ r8 = pix_d 5, 6
\r
1792 and r12,lr, r7, lsr #15
\r
1793 ldrh r12,[r3, r12]
\r
1794 and r7, lr, r7, lsr #23
\r
1796 and r12,r12,r9, lsl #2
\r
1798 mov r10,r10, lsr #1
\r
1799 sub r12,r12,r12,lsr #2 @ r12 = 3/4 pix_s 6
\r
1800 orr r10,r10,r12,lsl #16
\r
1801 and r7, r7, r9, lsl #2
\r
1802 add r10,r10,r7, lsl #14 @ r10 = pix_d 7, 8
\r
1806 stmia r0!, {r4,r5,r6,r8,r10}
\r
1808 bne .fl_loop32scale_RGB555u
\r
1812 ldmfd sp!, {r4-r10,lr}
\r
1815 #endif /* UNALIGNED_DRAWLINEDEST */
\r
1818 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1821 .global blockcpy @ void *dst, void *src, size_t n
\r
1824 stmfd sp!, {r4,r5}
\r
1825 mov r2, r2, lsr #4
\r
1827 ldmia r1!, {r3-r5,r12}
\r
1829 stmia r0!, {r3-r5,r12}
\r
1831 ldmfd sp!, {r4,r5}
\r
1835 .global blockcpy_or @ void *dst, void *src, size_t n, int pat
\r
1838 stmfd sp!, {r4-r6}
\r
1839 orr r3, r3, r3, lsl #8
\r
1840 orr r3, r3, r3, lsl #16
\r
1841 mov r2, r2, lsr #4
\r
1843 ldmia r1!, {r4-r6,r12}
\r
1849 stmia r0!, {r4-r6,r12}
\r
1850 bne blockcpy_loop_or
\r
1851 ldmfd sp!, {r4-r6}
\r
1854 @ vim:filetype=armasm
\r