1 @ vim:filetype=armasm
\r
3 @ ARM assembly versions of some funtions from draw.c
\r
4 @ this is highly specialized, be careful if changing related C code!
\r
6 @ (c) Copyright 2007-2008, Grazvydas "notaz" Ignotas
\r
7 @ All Rights Reserved
\r
9 .include "port_config.s"
\r
14 .extern DrawScanline
\r
17 .extern DrawLineDest
\r
18 .extern DrawStripInterlace
\r
19 .extern HighCacheS_ptr
\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 r4, [r1,#\ofs]
\r
160 orrlt r4, r3, r4 @ normal
\r
162 biceq r4, r4, #0xc0 @ hilight
\r
163 orreq r4, r4, #0x80
\r
164 orrgt r4, r4, #0xc0 @ shadow
\r
166 strb r4, [r1,#\ofs]
\r
170 @ TileFlipSh (r1=pdest, r2=pixels8, r3=pal) r4,r7: scratch, r0=sx, r12: register with helper pattern 0xf
\r
172 TileDoShGenPixel 16, 0 @ #0x000f0000
\r
173 TileDoShGenPixel 20, 1 @ #0x00f00000
\r
174 TileDoShGenPixel 24, 2 @ #0x0f000000
\r
175 TileDoShGenPixel 28, 3 @ #0xf0000000
\r
176 TileDoShGenPixel 0, 4 @ #0x0000000f
\r
177 TileDoShGenPixel 4, 5 @ #0x000000f0
\r
178 TileDoShGenPixel 8, 6 @ #0x00000f00
\r
179 TileDoShGenPixel 12, 7 @ #0x0000f000
\r
182 @ TileNormSh (r1=pdest, r2=pixels8, r3=pal) r4,r7: scratch, r0=sx, r12: register with helper pattern 0xf
\r
184 TileDoShGenPixel 12, 0 @ #0x0000f000
\r
185 TileDoShGenPixel 8, 1 @ #0x00000f00
\r
186 TileDoShGenPixel 4, 2 @ #0x000000f0
\r
187 TileDoShGenPixel 0, 3 @ #0x0000000f
\r
188 TileDoShGenPixel 28, 4 @ #0xf0000000
\r
189 TileDoShGenPixel 24, 5 @ #0x0f000000
\r
190 TileDoShGenPixel 20, 6 @ #0x00f00000
\r
191 TileDoShGenPixel 16, 7 @ #0x000f0000
\r
194 .macro TileDoShGenPixel_noop shift ofs
\r
196 and r4, r12, r2, lsr #\shift
\r
202 orrcc r4, r3, r4 @ 0-0xc (was 1-0xd)
\r
203 strccb r4, [r1,#\ofs]
\r
206 .macro TileFlipSh_noop
\r
207 TileDoShGenPixel_noop 16, 0 @ #0x000f0000
\r
208 TileDoShGenPixel_noop 20, 1 @ #0x00f00000
\r
209 TileDoShGenPixel_noop 24, 2 @ #0x0f000000
\r
210 TileDoShGenPixel_noop 28, 3 @ #0xf0000000
\r
211 TileDoShGenPixel_noop 0, 4 @ #0x0000000f
\r
212 TileDoShGenPixel_noop 4, 5 @ #0x000000f0
\r
213 TileDoShGenPixel_noop 8, 6 @ #0x00000f00
\r
214 TileDoShGenPixel_noop 12, 7 @ #0x0000f000
\r
217 .macro TileNormSh_noop
\r
218 TileDoShGenPixel_noop 12, 0 @ #0x0000f000
\r
219 TileDoShGenPixel_noop 8, 1 @ #0x00000f00
\r
220 TileDoShGenPixel_noop 4, 2 @ #0x000000f0
\r
221 TileDoShGenPixel_noop 0, 3 @ #0x0000000f
\r
222 TileDoShGenPixel_noop 28, 4 @ #0xf0000000
\r
223 TileDoShGenPixel_noop 24, 5 @ #0x0f000000
\r
224 TileDoShGenPixel_noop 20, 6 @ #0x00f00000
\r
225 TileDoShGenPixel_noop 16, 7 @ #0x000f0000
\r
228 .macro TileDoShGenPixel_onlyop_lp shift ofs
\r
230 ands r7, r12, r2, lsr #\shift
\r
234 ldrneb r4, [r1,#\ofs]
\r
239 biceq r4, r4, #0xc0 @ hilight
\r
240 orreq r4, r4, #0x80
\r
241 orrgt r4, r4, #0xc0 @ shadow
\r
242 strgeb 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
286 stmfd sp!, {r4-r11,lr}
\r
288 ldr r11, =(Pico+0x22228) @ Pico.video
\r
291 ldrb r7, [r11, #16] @ ??vv??hh
\r
293 mov r6, r1 @ hcache
\r
294 orr r9, r3, r0, lsl #30
\r
295 orr r9, r9, r2, lsl #8 @ r9=sh[31]|cellskip[15:8]|maxcells[7:0] (tmp)
\r
298 orr r1, r1, #0x00ff
\r
302 biclt r1, r1, #0xfc00
\r
303 biceq r1, r1, #0xfe00
\r
304 bicgt r1, r1, #0xff00 @ r1=ymask=(height<<8)|0xff; ...; // Y Mask in pixels
\r
308 subge r10, r10, #1 @ r10=shift[width] (5,6,6,7)
\r
311 mov r5, r8, lsl r10
\r
312 sub r5, r5, #1 @ r5=xmask
\r
316 ldreqb r12, [r11, #2]
\r
317 ldrneb r12, [r11, #4]
\r
319 ldr r2, =DrawScanline @ trying to make good use of pipeline here
\r
320 ldr lr, =(Pico+0x10000) @ lr=Pico.vram
\r
322 moveq r12, r12, lsl #10
\r
323 movne r12, r12, lsl #13
\r
324 and r12, r12, #(7<<13) @ r12=(ts->nametab<<1) (halfword compliant)
\r
326 ldrh r8, [r11, #12]
\r
327 ldrb r7, [r11, #11]
\r
330 mov r4, r8, lsr #8 @ pvid->reg[13]
\r
331 mov r4, r4, lsl #10 @ htab=pvid->reg[13]<<9; (halfwords)
\r
333 addne r4, r4, r2, lsl #2 @ htab+=DrawScanline<<1; // Offset by line
\r
335 biceq r4, r4, #0x1f @ htab&=~0xf; // Offset by tile
\r
336 add r4, r4, r0, lsl #1 @ htab+=plane
\r
337 bic r4, r4, #0x00ff0000 @ just in case
\r
338 ldrh r3, [lr, r4] @ r3=hscroll
\r
341 bne .DrawStrip_vsscroll
\r
343 @ Get vertical scroll value:
\r
344 add r7, lr, #0x012000
\r
345 add r7, r7, #0x000180 @ r7=Pico.vsram (Pico+0x22180)
\r
350 bne .DrawStrip_interlace
\r
353 movne r7, r7, lsr #16
\r
355 @ Find the line in the name table
\r
359 add r10, r10, #1 @ shift[width]++
\r
360 add r12, r12, r4, lsl r10 @ nametab+=(ts.line>>3)<<shift[width];
\r
362 @ ldmia r0, {r1,r2,r3,r5,r6,r9} @ r2=line, r3=ts->hscroll, r5=ts->xmask, r6=ts->hc, r9=ts->cells
\r
365 mov r10,r10, lsl #1 @ r10=ty=(ts->line&7)<<1;
\r
366 orr r10,r10, r9, lsl #24
\r
369 mov r8, r8, lsr #3 @ r8=tilex=(-ts->hscroll)>>3
\r
373 add r7, r1, #1 @ r7=dx=((ts->hscroll-1)&7)+1
\r
377 orrne r10,r10, #1<<23 @ r10=(cells<<24|sh<<23|hi_not_empty<<22|had_output<<21|ty)
\r
378 movne r3, #0x40 @ default to shadowed pal on sh mode
\r
381 addne r10,r10, #0x01000000 @ we will loop cells+1 times if there is scroll
\r
383 and r9, r9, #0xff00
\r
384 add r8, r8, r9, lsr #8 @ tilex+=cellskip
\r
385 add r7, r7, r9, lsr #5 @ dx+=cellskip<<3;
\r
386 sub r10,r10,r9, lsl #16 @ cells-=cellskip
\r
388 @ cache some stuff to avoid mem access
\r
389 .if OVERRIDE_HIGHCOL
\r
398 mvn r9, #0 @ r9=prevcode=-1
\r
399 add r1, r11, r7 @ r1=pdest
\r
402 @ r4 & r7 are scratch in this loop
\r
405 .dsloop: @ 40-41 times
\r
406 subs r10,r10, #0x01000000
\r
411 add r7, lr, r7, lsl #1 @ Pico.vram+((tilex&ts->xmask) as halfwords)
\r
412 ldrh r7, [r7, r12] @ r7=code (int, but from unsigned, no sign extend)
\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, #0x00c00000
\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 ldreq r1, =rendstatus
\r
501 str r0, [r6] @ terminate the cache list
\r
502 orreq r2, r2, #0x40 @ had a layer with all hi-prio tiles
\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 r4, =DrawScanline
\r
517 orr r5, r1, r10, lsl #24
\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 .if OVERRIDE_HIGHCOL
\r
551 mvn r9, #0 @ r9=prevcode=-1
\r
552 add r1, r11, r7 @ r1=pdest
\r
554 @ r4 & r7 are scratch in this loop
\r
557 .dsloop_vs: @ 40-41 times
\r
558 add r10,r10, #0x01000000
\r
559 and r4, r10, #0x003f0000
\r
560 cmp r4, r10, asr #8
\r
561 ble .dsloop_vs_exit
\r
563 @ calc offset and read tileline code to r7, also calc ty
\r
564 add r7, lr, #0x012000
\r
565 add r7, r7, #0x000180 @ r7=Pico.vsram (Pico+0x22180)
\r
566 add r7, r7, r10,asr #23 @ vsram + ((cell&~1)<<1)
\r
568 tst r10,#0x8000 @ plane1?
\r
570 ldrh r7, [r7] @ r7=vscroll
\r
572 bic r10,r10,#0xff @ clear old ty
\r
573 and r4, r5, #0xff0000 @ scanline
\r
574 add r4, r4, r7, lsl #16 @ ... += vscroll
\r
575 and r4, r4, r5, lsl #16 @ ... &= ymask
\r
576 and r7, r4, #0x70000
\r
577 orr r10,r10,r7, lsr #15 @ new ty
\r
579 mov r4, r4, lsr #19
\r
580 mov r7, r5, lsr #24
\r
581 mov r4, r4, lsl r7 @ nametabadd
\r
583 and r7, r8, r8, lsr #25
\r
584 add r7, lr, r7, lsl #1 @ Pico.vram+((tilex&ts->xmask) as halfwords)
\r
585 add r7, r7, r4, lsl #1
\r
586 ldrh r7, [r7, r12] @ r7=code (int, but from unsigned, no sign extend)
\r
592 bne .DrawStrip_vs_hiprio
\r
595 beq .DrawStrip_vs_samecode @ we know stuff about this tile already
\r
597 mov r9, r7 @ remember code
\r
598 orr r8, r8, #(1<<24)@ seen non hi-prio tile
\r
600 movs r2, r9, lsl #20 @ if (code&0x1000)
\r
602 add r2, r2, r10, lsl #17
\r
603 mov r2, r2, lsr #17
\r
604 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
606 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
609 and r3, r9, #0x6000
\r
610 add r3, r7, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);
\r
612 .DrawStrip_vs_samecode:
\r
614 beq .dsloop_vs @ tileline blank
\r
617 beq .DrawStrip_vs_SingleColor @ tileline singlecolor
\r
620 bne .DrawStrip_vs_TileFlip
\r
622 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r0: helper pattern
\r
623 .DrawStrip_vs_TileNorm:
\r
627 .DrawStrip_vs_TileFlip:
\r
631 .DrawStrip_vs_SingleColor:
\r
634 orr r4, r4, r4, lsl #8
\r
635 tst r1, #1 @ not aligned?
\r
636 strneb r4, [r1], #1
\r
637 streqh r4, [r1], #2
\r
641 strneb r4, [r1], #1 @ have a remaining unaligned pixel?
\r
644 .DrawStrip_vs_hiprio:
\r
645 tst r10, #0x00c00000
\r
646 beq .DrawStrip_vs_hiprio_maybempt
\r
648 orr r7, r7, r0, lsl #16
\r
649 orr r7, r7, r10, lsl #25 @ (ty<<25)
\r
651 eorne r7, r7, #7<<26 @ if(code&0x1000) cval^=7<<26;
\r
652 str r7, [r6], #4 @ cache hi priority tile
\r
656 .DrawStrip_vs_hiprio_maybempt:
\r
658 beq .dsloop_vs @ must've been empty, otherwise we wouldn't get here
\r
659 movs r2, r7, lsl #20 @ if (code&0x1000)
\r
661 add r2, r2, r10, lsl #17
\r
662 mov r2, r2, lsr #17
\r
663 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
664 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
665 mov r9, r7 @ remember code
\r
667 orrne r10, r10, #1<<22
\r
668 bne .DrawStrip_vs_hiprio
\r
672 tst r8, #(1<<24) @ seen non hi-prio tile
\r
673 ldreq r1, =rendstatus
\r
676 str r0, [r6] @ terminate the cache list
\r
677 orreq r2, r2, #0x40 @ had a layer with all hi-prio tiles
\r
680 ldmfd sp!, {r4-r11,lr}
\r
684 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
686 @ interlace mode 2? Sonic 2?
\r
687 .DrawStrip_interlace:
\r
689 moveq r7, r7, lsl #21
\r
690 movne r7, r7, lsl #5
\r
692 @ Find the line in the name table
\r
693 add r2, r7, r2, lsl #22 @ r2=(vscroll+(DrawScanline<<1))<<21 (11 bits);
\r
694 orr r1, r1, #0x80000000
\r
695 and r2, r2, r1, ror #10 @ &((ymask<<1)|1)<<21;
\r
696 mov r2, r2, lsr #21
\r
698 mov r12, r12, lsr #1 @ halfwords
\r
699 add r0, r12, r4, lsl r10 @ nametab+=(ts.line>>4)<<shift[width];
\r
703 stmia sp, {r0,r2,r3,r5,r6,r9}
\r
706 bl DrawStripInterlace @ struct TileStrip *ts
\r
709 ldmfd sp!, {r4-r11,lr}
\r
714 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
717 .global BackFill @ int reg7, int sh
\r
720 stmfd sp!, {r4-r9,lr}
\r
722 .if OVERRIDE_HIGHCOL
\r
724 mov r0, r0, lsl #26
\r
726 mov r0, r0, lsr #26
\r
729 ldr lr, =(HighCol+8)
\r
730 mov r0, r0, lsl #26
\r
731 mov r0, r0, lsr #26
\r
734 orr r0, r0, r1, lsl #6
\r
735 orr r0, r0, r0, lsl #8
\r
736 orr r0, r0, r0, lsl #16
\r
747 stmia lr!, {r0-r7} @ 10*8*4
\r
758 ldmfd sp!, {r4-r9,r12}
\r
762 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
765 .global DrawTilesFromCache @ int *hc, int sh, int rlim
\r
767 DrawTilesFromCache:
\r
768 stmfd sp!, {r4-r8,r11,lr}
\r
770 @ cache some stuff to avoid mem access
\r
771 .if OVERRIDE_HIGHCOL
\r
779 ldr lr, =(Pico+0x10000) @ lr=Pico.vram
\r
781 mvn r5, #0 @ r5=prevcode=-1
\r
783 orr r8, r8, r2, lsl #1
\r
784 bne .dtfc_check_rendflags
\r
788 ldr r6, [r0], #4 @ read code
\r
789 movs r1, r6, lsr #16 @ r1=dx;
\r
790 ldmeqfd sp!, {r4-r8,r11,pc} @ dx is never zero, this must be a terminator, return
\r
791 bic r4, r1, #0xfe00
\r
792 add r1, r11, r4 @ r1=pdest
\r
794 mov r7, r6, lsl #16
\r
795 cmp r5, r7, lsr #16
\r
796 beq .dtfc_samecode @ if (code==prevcode)
\r
798 mov r5, r7, lsr #16
\r
800 mov r2, r5, lsl #21
\r
801 mov r2, r2, lsr #17 @ r2=addr=(code&0x7ff)<<4;
\r
802 add r2, r2, r6, lsr #25 @ addr+=ty
\r
804 and r3, r5, #0x6000
\r
805 mov r3, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);
\r
807 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
810 rsbs r4, r4, r8, lsr #1
\r
820 beq .dtfc_SingleColor @ tileline singlecolor
\r
825 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
837 orr r4, r4, r4, lsl #8
\r
838 tst r1, #1 @ not aligned?
\r
839 strneb r4, [r1], #1
\r
840 streqh r4, [r1], #2
\r
844 strneb r4, [r1], #1 @ have a remaining unaligned pixel?
\r
849 beq .dtfc_shadow_blank
\r
852 beq .dtfc_SingleColor @ tileline singlecolor
\r
855 bne .dtfc_TileFlipShHP
\r
857 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
858 .dtfc_TileNormShHP:
\r
862 .dtfc_TileFlipShHP:
\r
866 .dtfc_shadow_blank:
\r
871 strneb r4, [r1], #1
\r
873 orr r6, r6, r6, lsl #8
\r
889 add r4, r4, #7 @ 0-6
\r
893 mov r2, r2, ror #16
\r
894 tst r5, #0x0800 @ flipped?
\r
897 mov r2, r2, ror #16
\r
907 @ check if we have detected layer covered with hi-prio tiles:
\r
908 .dtfc_check_rendflags:
\r
909 ldr r1, =rendstatus
\r
913 bic r8, r8, #1 @ sh/hi mode off
\r
915 bne .dtfc_loop @ already processed
\r
922 orr r6, r6, r6, lsl #8
\r
923 orr r6, r6, r6, lsl #16
\r
925 ldmia r1, {r2,r4,r5,r7}
\r
931 stmia r1!,{r2,r4,r5,r7}
\r
932 bne .dtfc_loop_shprep
\r
934 mvn r5, #0 @ r5=prevcode=-1
\r
939 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
942 .global DrawSpritesFromCache @ int *hc, int maxwidth, int prio, int sh
\r
944 DrawSpritesFromCache:
\r
945 stmfd sp!, {r4-r11,lr}
\r
947 @ cache some stuff to avoid mem access
\r
948 .if OVERRIDE_HIGHCOL
\r
956 ldr lr, =(Pico+0x10000) @ lr=Pico.vram
\r
957 mov r6, r3, lsl #31
\r
963 ldr r9, [r10], #4 @ read code
\r
964 bic r6, r6, #7 @ using pipeline
\r
966 ldmeqfd sp!, {r4-r11,pc}
\r
968 mov r4, r9, lsl #28
\r
969 orr r6, r6, r4, lsr #30
\r
970 add r6, r6, #1 @ r6=s1cc???? ... ?????www (s=shadow/hilight, cc=pal, w=width)
\r
973 add r5, r5, #1 @ r5=delta
\r
975 rsbne r5, r5, #0 @ Flip X
\r
978 mov r2, r9, lsr #17
\r
979 mov r8, r2, lsl #1 @ tile=((unsigned int)code>>17)<<1;
\r
981 and r3, r9, #0x30 @ r3=pal=(code&0x30);
\r
984 orr r6, r6, r3, lsl #24
\r
986 mov r0, r9, lsl #16
\r
987 mov r0, r0, asr #22 @ sx=(code<<16)>>22
\r
988 adds r0, r0, #0 @ set ZV
\r
989 b .dsfc_inloop_enter
\r
999 .dsfc_inloop_enter:
\r
1004 mov r8, r8, lsl #17
\r
1005 mov r8, r8, lsr #17 @ tile&=0x7fff; // Clip tile address
\r
1007 ldr r2, [lr, r8, lsl #1] @ pack=*(unsigned int *)(Pico.vram+tile); // Get 8 pixels
\r
1008 add r1, r11, r0 @ r1=pdest
\r
1012 cmp r12, r6, lsr #28
\r
1015 cmp r2, r2, ror #4
\r
1016 beq .dsfc_SingleColor @ tileline singlecolor
\r
1019 bne .dsfc_TileFlip
\r
1021 @ TileFlip (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
1030 .dsfc_SingleColor:
\r
1031 tst r0, #1 @ not aligned?
\r
1034 orr r4, r4, r4, lsl #8
\r
1035 strneb r4, [r1], #1
\r
1036 streqh r4, [r1], #2
\r
1040 strneb r4, [r1], #1
\r
1044 tst r9, #0x80000000
\r
1045 beq .dsfc_shadow_lowpri
\r
1047 cmp r2, r2, ror #4
\r
1048 beq .dsfc_singlec_sh
\r
1051 bne .dsfc_TileFlip_sh
\r
1053 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
1054 .dsfc_TileNorm_sh:
\r
1058 .dsfc_TileFlip_sh:
\r
1063 cmp r2, #0xe0000000
\r
1064 bcc .dsfc_SingleColor @ normal singlecolor tileline (carry inverted in ARM)
\r
1065 tst r2, #0x10000000
\r
1074 .dsfc_shadow_lowpri:
\r
1076 bne .dsfc_TileFlip_sh_lp
\r
1078 .dsfc_TileNorm_sh_lp:
\r
1079 TileNormSh_onlyop_lp
\r
1082 .dsfc_TileFlip_sh_lp:
\r
1083 TileFlipSh_onlyop_lp
\r
1088 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1090 @ + 0 : hhhhvvvv ab--hhvv yyyyyyyy yyyyyyyy // a: offscreen h, b: offs. v, h: horiz. size
\r
1091 @ + 4 : xxxxxxxx xxxxxxxx pccvhnnn nnnnnnnn // x: x coord + 8
\r
1093 .global DrawSprite @ unsigned int *sprite, int sh, int acc_sprites
\r
1096 stmfd sp!, {r4-r9,r11,lr}
\r
1098 orr r8, r2, r1, lsl #4
\r
1099 ldr r3, [r0] @ sprite[0]
\r
1100 ldr r7, =DrawScanline
\r
1101 mov r6, r3, lsr #28
\r
1102 sub r6, r6, #1 @ r6=width-1 (inc later)
\r
1103 mov r5, r3, lsr #24
\r
1104 and r5, r5, #7 @ r5=height
\r
1106 mov r4, r3, lsl #16 @ r4=sy<<16 (tmp)
\r
1110 sub r7, r7, r4, asr #16 @ r7=row=DrawScanline-sy
\r
1112 mov r2, r9, asr #16 @ r2=sx
\r
1113 mov r9, r9, lsl #16
\r
1114 mov r9, r9, lsr #16
\r
1115 orr r9, r9, r8, lsl #27 @ r9=code|sh[31]|as[27]
\r
1118 movne r4, r5, lsl #3
\r
1120 subne r7, r4, r7 @ if (code&0x1000) row=(height<<3)-1-row; // Flip Y
\r
1122 add r8, r9, r7, lsr #3 @ tile+=row>>3; // Tile number increases going down
\r
1124 mlane r8, r5, r6, r8 @ if (code&0x0800) { tile+=delta*(width-1);
\r
1125 rsbne r5, r5, #0 @ delta=-delta; } // r5=delta now
\r
1127 mov r8, r8, lsl #21
\r
1128 mov r8, r8, lsr #17
\r
1130 add r8, r8, r7, lsl #1 @ tile+=(row&7)<<1; // Tile address
\r
1133 tsteq r9, #(1<<27)
\r
1134 bne .dspr_cache @ if(code&0x8000) || as
\r
1137 tstne r9, #(1<<31)
\r
1138 bne .dspr_cache @ (sh && pal == 0x30)
\r
1141 @ cache some stuff to avoid mem access
\r
1142 .if OVERRIDE_HIGHCOL
\r
1150 ldr lr, =(Pico+0x10000) @ lr=Pico.vram
\r
1152 mov r5, r5, lsl #4 @ delta<<=4; // Delta of address
\r
1153 and r4, r9, #0x6000
\r
1154 orr r9, r9, r4, lsl #16
\r
1155 orrs r9, r9, #0x10000000 @ r9=scc1 a??? ... <code> (s=shadow/hilight, cc=pal, a=acc_spr)
\r
1157 mov r3, r4, lsr #9 @ r3=pal=((code>>9)&0x30);
\r
1158 orrmi r3, r3, #0x40 @ shadow by default
\r
1160 add r6, r6, #1 @ inc now
\r
1161 adds r0, r2, #0 @ mov sx to r0 and set ZV flags
\r
1162 b .dspr_loop_enter
\r
1165 subs r6, r6, #1 @ width--
\r
1166 ldmeqfd sp!, {r4-r9,r11,pc}@ return
\r
1167 adds r0, r0, #8 @ sx+=8
\r
1168 add r8, r8, r5 @ tile+=delta
\r
1171 ble .dspr_loop @ sx <= 0
\r
1173 ldmgefd sp!, {r4-r9,r11,pc}@ return
\r
1175 mov r8, r8, lsl #17
\r
1176 mov r8, r8, lsr #17 @ tile&=0x7fff; // Clip tile address
\r
1178 ldr r2, [lr, r8, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
1179 add r1, r11, r0 @ r1=pdest
\r
1183 cmp r12, r9, lsr #28
\r
1186 cmp r2, r2, ror #4
\r
1187 beq .dspr_SingleColor @ tileline singlecolor
\r
1190 bne .dspr_TileFlip
\r
1192 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
1203 cmp r2, #0xe0000000
\r
1204 bcs .dspr_loop @ operator tileline, ignore
\r
1206 .dspr_SingleColor:
\r
1209 orr r4, r4, r4, lsl #8
\r
1210 tst r0, #1 @ not aligned?
\r
1211 strneb r4, [r1], #1
\r
1212 streqh r4, [r1], #2
\r
1216 strneb r4, [r1], #1
\r
1220 cmp r2, r2, ror #4
\r
1221 beq .dspr_singlec_sh
\r
1224 bne .dspr_TileFlip_sh
\r
1226 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
1227 .dspr_TileNorm_sh:
\r
1231 .dspr_TileFlip_sh:
\r
1237 @ *HighCacheS_ptr++ = ((code&0x8000)<<16)|(tile<<16)|((code&0x0800)<<5)|((sx<<6)&0x0000ffc0)|pal|((sprite[0]>>16)&0xf);
\r
1238 ldr r1, =HighCacheS_ptr
\r
1239 mov r4, r8, lsl #16 @ tile
\r
1241 orrne r4, r4, #0x10000 @ code&0x0800
\r
1242 mov r0, r2, lsl #22
\r
1243 orr r4, r4, r0, lsr #16 @ (sx<<6)&0x0000ffc0
\r
1244 and r0, r9, #0x6000
\r
1245 orr r4, r4, r0, lsr #9 @ (code>>9)&0x30
\r
1246 mov r3, r3, lsl #12
\r
1247 orr r4, r4, r3, lsr #28 @ (sprite[0]>>24)&0xf
\r
1251 orrne r4, r4, #0x80000000 @ prio
\r
1256 and r0, r9, #(1<<27) @ as
\r
1257 teqne r0, #(1<<27) @ (code&0x8000) && !as
\r
1258 ldmnefd sp!, {r4-r9,r11,pc}
\r
1259 b .dspr_continue @ draw anyway if accurate sprites enabled
\r
1261 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1263 .global DrawWindow @ int tstart, int tend, int prio, int sh // int *hcache
\r
1266 stmfd sp!, {r4-r11,lr}
\r
1268 ldr r11, =(Pico+0x22228) @ Pico.video
\r
1269 ldr r10, =DrawScanline
\r
1270 ldrb r12, [r11, #3] @ pvid->reg[3]
\r
1273 ldr r4, [r11, #12]
\r
1274 mov r5, r10, lsr #3
\r
1276 mov r10, r10, lsl #1 @ r10=ty
\r
1278 mov r12, r12, lsl #10
\r
1280 tst r4, #1 @ 40 cell mode?
\r
1281 andne r12, r12, #0xf000 @ 0x3c<<10
\r
1282 andeq r12, r12, #0xf800
\r
1283 addne r12, r12, r5, lsl #7
\r
1284 addeq r12, r12, r5, lsl #6 @ nametab
\r
1285 add r12, r12, r0, lsl #2 @ +starttile
\r
1287 ldr r6, =rendstatus
\r
1288 ldr lr, =(Pico+0x10000) @ lr=Pico.vram
\r
1291 @ fetch the first code now
\r
1292 ldrh r7, [lr, r12]
\r
1294 ands r6, r6, #2 @ we care about bit 1 only
\r
1297 eoreq r8, r2, r7, lsr #15 @ do prio bits differ?
\r
1299 ldmeqfd sp!, {r4-r11,pc} @ yes, assume that whole window uses same priority
\r
1301 orr r6, r6, r3, lsl #8 @ shadow mode
\r
1305 @ cache some stuff to avoid mem access
\r
1306 .if OVERRIDE_HIGHCOL
\r
1308 mov r8, r8, lsl #1 @ cells
\r
1310 mvn r9, #0 @ r9=prevcode=-1
\r
1313 ldr r11,=(HighCol+8)
\r
1314 mov r8, r8, lsl #1 @ cells
\r
1315 mvn r9, #0 @ r9=prevcode=-1
\r
1317 add r1, r11, r0, lsl #4 @ r1=pdest
\r
1321 @ r4,r5 are scratch in this loop
\r
1325 add r12, r12, #2 @ halfwords
\r
1326 ldrh r7, [lr, r12] @ r7=code (int, but from unsigned, no sign extend)
\r
1328 beq .dwloop_end @ done
\r
1330 eor r5, r6, r7, lsr #15
\r
1332 orrne r6, r6, #2 @ wrong pri
\r
1336 beq .dw_samecode @ we know stuff about this tile already
\r
1339 mov r9, r7 @ remember code
\r
1341 movs r2, r9, lsl #20 @ if (code&0x1000)
\r
1342 mov r2, r2, lsl #1
\r
1343 add r2, r10, r2, lsr #17 @ r2=addr=(code&0x7ff)<<4; addr+=ty
\r
1344 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
1346 and r3, r9, #0x6000
\r
1347 mov r3, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);
\r
1349 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
1356 beq .dwloop @ tileline blank
\r
1358 cmp r2, r2, ror #4
\r
1359 beq .dw_SingleColor @ tileline singlecolor
\r
1364 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r0: helper pattern
\r
1374 and r4, r0, r2 @ #0x0000000f
\r
1376 orr r4, r4, r4, lsl #8
\r
1377 orr r4, r4, r4, lsl #16
\r
1379 stmia r1!, {r4,r5}
\r
1380 b .dwloop_nor1 @ we incremeted r1 ourselves
\r
1383 tst r6, #1 @ hi pri?
\r
1384 orreq r3, r3, #0x40
\r
1385 beq .dw_shadow_done
\r
1388 orr r5, r5, r5, lsl #8
\r
1389 orr r5, r5, r5, lsl #16
\r
1398 ldr r0, =rendstatus
\r
1404 ldmfd sp!, {r4-r11,r12}
\r
1408 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1411 @ hilights 2 pixels in RGB444/BGR444 format
\r
1412 .macro TileDoShHi2Pixels444 reg
\r
1413 mov \reg, \reg, ror #12
\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 #24
\r
1423 adds \reg, \reg, #0x40000000
\r
1424 orrcs \reg, \reg, #0xf0000000
\r
1425 mov \reg, \reg, ror #28
\r
1426 adds \reg, \reg, #0x40000000
\r
1427 orrcs \reg, \reg, #0xf0000000
\r
1428 mov \reg, \reg, ror #28
\r
1429 adds \reg, \reg, #0x40000000
\r
1430 orrcs \reg, \reg, #0xf0000000
\r
1431 mov \reg, \reg, ror #12
\r
1435 .global FinalizeLineBGR444 @ int sh
\r
1437 FinalizeLineBGR444:
\r
1438 stmfd sp!, {r4-r6,lr}
\r
1440 ldr lr, =(Pico+0x22228) @ Pico.video
\r
1441 ldr r0, =DrawLineDest
\r
1442 ldrb r12, [lr, #12]
\r
1444 sub r3, lr, #0x128 @ r3=Pico.cram
\r
1447 movne r2, #320/4 @ len
\r
1448 bne .fl_no32colBGR444
\r
1453 addeq r0, r0, #32*2
\r
1455 .fl_no32colBGR444:
\r
1457 beq .fl_noshBGR444
\r
1461 ldrb r12, [lr, #-0x1a] @ 0x2220e ~ dirtyPal
\r
1464 beq .fl_noshBGR444
\r
1466 strb r12, [lr, #-0x1a]
\r
1471 ldmia r3!, {r1,r5,r6,r12}
\r
1473 stmia r4!, {r1,r5,r6,r12}
\r
1474 bne .fl_loopcpBGR444
\r
1476 @ shadowed pixels:
\r
1478 orr r12,r12,#0x0700
\r
1479 orr r12,r12,r12,lsl #16
\r
1480 sub r3, r3, #0x40*2
\r
1481 add r5, r4, #0x80*2
\r
1483 .fl_loopcpBGR444_sh:
\r
1484 ldmia r3!, {r1,r6}
\r
1486 and r1, r12, r1, lsr #1
\r
1487 and r6, r12, r6, lsr #1
\r
1488 stmia r4!, {r1,r6}
\r
1489 stmia r5!, {r1,r6}
\r
1490 bne .fl_loopcpBGR444_sh
\r
1492 @ hilighted pixels:
\r
1493 sub r3, r3, #0x40*2
\r
1495 .fl_loopcpBGR444_hi:
\r
1497 TileDoShHi2Pixels444 r1
\r
1500 bne .fl_loopcpBGR444_hi
\r
1502 sub r3, r4, #0x40*3*2
\r
1507 ldr r12,=rendstatus
\r
1508 eors r6, r6, #1 @ sh is 0
\r
1511 tstne r12,#(1<<2) @ and PDRAW_ACC_SPRITES
\r
1513 .if OVERRIDE_HIGHCOL
\r
1517 mov lr, lr, lsl #1
\r
1520 ldr r1, =(HighCol+8)
\r
1522 mov lr, lr, lsl #1
\r
1529 and r4, lr, r12, lsl #1
\r
1531 and r5, lr, r12, lsr #7
\r
1533 and r6, lr, r12, lsr #15
\r
1535 and r12,lr, r12, lsr #23
\r
1536 ldrh r12,[r3, r12] @ 1c.i.
\r
1537 orr r4, r4, r5, lsl #16
\r
1538 orr r5, r6, r12,lsl #16
\r
1540 stmia r0!, {r4,r5}
\r
1541 bne .fl_loopBGR444
\r
1544 ldmfd sp!, {r4-r6,lr}
\r
1548 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1551 @ hilights 2 pixels in RGB555/BGR555 format
\r
1552 .macro TileDoShHi2Pixels555 reg
\r
1553 adds \reg, \reg, #0x40000000
\r
1554 orrcs \reg, \reg, #0xf0000000
\r
1555 mov \reg, \reg, ror #27
\r
1556 adds \reg, \reg, #0x40000000
\r
1557 orrcs \reg, \reg, #0xf0000000
\r
1558 mov \reg, \reg, ror #26
\r
1559 adds \reg, \reg, #0x40000000
\r
1560 orrcs \reg, \reg, #0xf0000000
\r
1561 mov \reg, \reg, ror #27
\r
1562 adds \reg, \reg, #0x40000000
\r
1563 orrcs \reg, \reg, #0xf0000000
\r
1564 mov \reg, \reg, ror #27
\r
1565 adds \reg, \reg, #0x40000000
\r
1566 orrcs \reg, \reg, #0xf0000000
\r
1567 mov \reg, \reg, ror #26
\r
1568 adds \reg, \reg, #0x40000000
\r
1569 orrcs \reg, \reg, #0xf0000000
\r
1570 mov \reg, \reg, ror #27
\r
1574 @ Convert 0000bbb0 ggg0rrr0
\r
1575 @ to rrrrrggg gggbbbbb
\r
1577 @ r2,r3,r9 - scratch, lr = 0x001c001c, r8 = 0x00030003
\r
1578 .macro convRGB565 reg
\r
1579 and r2, lr, \reg,lsl #1
\r
1580 and r9, r8, \reg,lsr #2
\r
1581 orr r2, r2, r9 @ r2=red
\r
1582 and r3, lr, \reg,lsr #7
\r
1583 and r9, r8, \reg,lsr #10
\r
1584 orr r3, r3, r9 @ r3=blue
\r
1585 and \reg, \reg, lr, lsl #3
\r
1586 orr \reg, \reg, \reg,lsl #3 @ green
\r
1587 orr \reg, \reg, r2, lsl #11 @ add red back
\r
1588 orr \reg, \reg, r3 @ add blue back
\r
1591 .global vidConvCpyRGB565
\r
1593 vidConvCpyRGB565: @ void *to, void *from, int pixels
\r
1594 stmfd sp!, {r4-r9,lr}
\r
1596 mov r12, r2, lsr #3 @ repeats
\r
1597 mov lr, #0x001c0000
\r
1598 orr lr, lr, #0x01c @ lr == pattern 0x001c001c
\r
1599 mov r8, #0x00030000
\r
1600 orr r8, r8, #0x003
\r
1603 ldmia r1!, {r4-r7}
\r
1616 ldmfd sp!, {r4-r9,lr}
\r
1621 .global FinalizeLineRGB555 @ int sh
\r
1623 FinalizeLineRGB555:
\r
1624 stmfd sp!, {r4-r8,lr}
\r
1625 ldr r8, =(Pico+0x22228) @ Pico.video
\r
1628 ldrb r7, [r8, #-0x1a] @ 0x2220e ~ dirtyPal
\r
1632 beq .fl_noconvRGB555
\r
1633 strb r1, [r8, #-0x1a]
\r
1634 sub r1, r8, #0x128 @ r1=Pico.cram
\r
1637 bl vidConvCpyRGB565
\r
1642 beq .fl_noshRGB555
\r
1644 beq .fl_noshRGB555
\r
1646 @ shadowed pixels:
\r
1648 orr r12,r12,#0x7300
\r
1649 orr r12,r12,r12,lsl #16
\r
1650 add r4, r3, #0x40*2
\r
1651 add r5, r3, #0xc0*2
\r
1653 .fl_loopcpRGB555_sh:
\r
1654 ldmia r3!, {r1,r6}
\r
1656 and r1, r12, r1, lsr #1
\r
1657 and r6, r12, r6, lsr #1
\r
1658 stmia r4!, {r1,r6}
\r
1659 stmia r5!, {r1,r6}
\r
1660 bne .fl_loopcpRGB555_sh
\r
1662 @ hilighted pixels:
\r
1663 sub r3, r3, #0x40*2
\r
1665 .fl_loopcpRGB555_hi:
\r
1667 TileDoShHi2Pixels555 r1
\r
1670 bne .fl_loopcpRGB555_hi
\r
1672 sub r3, r3, #0x40*2
\r
1676 ldr r12,=rendstatus
\r
1677 eors r6, r6, #1 @ sh is 0
\r
1680 tstne r12,#(1<<2) @ and PDRAW_ACC_SPRITES
\r
1683 .if OVERRIDE_HIGHCOL
\r
1685 ldr r0, =DrawLineDest
\r
1690 ldr r0, =DrawLineDest
\r
1691 ldr r1, =(HighCol+8)
\r
1695 ldrb r12, [r8, #12]
\r
1696 mov lr, lr, lsl #1
\r
1699 movne r2, #320/8 @ len
\r
1700 bne .fl_no32colRGB555
\r
1705 bne .fl_32scale_RGB555
\r
1707 addeq r0, r0, #32*2
\r
1709 .fl_no32colRGB555:
\r
1711 .if UNALIGNED_DRAWLINEDEST
\r
1712 @ this is basically for Gizmondo, which has unaligned odd lines in the framebuffer
\r
1721 and r4, lr, r12, lsl #1
\r
1723 and r5, lr, r12, lsr #7
\r
1725 and r6, lr, r12, lsr #15
\r
1727 orr r4, r4, r5, lsl #16
\r
1729 and r5, lr, r12, lsr #23
\r
1731 and r8, lr, r7, lsl #1
\r
1733 orr r5, r6, r5, lsl #16
\r
1735 and r6, lr, r7, lsr #7
\r
1737 and r12,lr, r7, lsr #15
\r
1738 ldrh r12,[r3, r12]
\r
1739 and r7, lr, r7, lsr #23
\r
1741 orr r8, r8, r6, lsl #16
\r
1744 orr r12,r12, r7, lsl #16
\r
1746 stmia r0!, {r4,r5,r8,r12}
\r
1747 bne .fl_loopRGB555
\r
1749 ldmfd sp!, {r4-r8,lr}
\r
1753 .fl_32scale_RGB555:
\r
1754 stmfd sp!, {r9,r10}
\r
1755 mov r9, #0x3900 @ f800 07e0 001f | e000 0780 001c | 3800 01e0 0007
\r
1756 orr r9, r9, #0x00e7
\r
1758 .if UNALIGNED_DRAWLINEDEST
\r
1760 bne .fl_32scale_RGB555u
\r
1763 .fl_loop32scale_RGB555:
\r
1767 and r4, lr, r12,lsl #1
\r
1769 and r5, lr, r12,lsr #7
\r
1771 and r4, r4, r9, lsl #2
\r
1772 orr r4, r4, r4, lsl #14 @ r4[31:16] = 1/4 pix_s 0
\r
1773 and r5, r5, r9, lsl #2
\r
1774 sub r6, r5, r5, lsr #2 @ r6 = 3/4 pix_s 1
\r
1775 add r4, r4, r6, lsl #16 @ pix_d 0, 1
\r
1776 and r6, lr, r12,lsr #15
\r
1778 and r12,lr, r12,lsr #23
\r
1779 ldrh r12,[r3, r12]
\r
1780 and r6, r6, r9, lsl #2
\r
1782 mov r5, r5, lsr #1
\r
1783 sub r6, r6, r6, lsr #2 @ r6 = 3/4 pix_s 2
\r
1784 orr r5, r5, r6, lsl #16
\r
1786 and r6, lr, r7, lsl #1
\r
1788 and r12,r12,r9, lsl #2
\r
1789 add r5, r5, r12,lsl #14 @ pix_d 2, 3
\r
1790 and r6, r6, r9, lsl #2
\r
1791 orr r6, r12,r6, lsl #16 @ pix_d 4, 5
\r
1793 and r12,lr, r7, lsr #7
\r
1794 ldrh r12,[r3, r12]
\r
1795 and r10,lr, r7, lsr #15
\r
1796 ldrh r10,[r3, r10]
\r
1797 and r12,r12,r9, lsl #2
\r
1798 sub r8, r12,r12,lsr #2 @ r8 = 3/4 pix_s 1
\r
1799 add r8, r8, r6, lsr #18
\r
1800 and r7, lr, r7, lsr #23
\r
1802 and r10,r10,r9, lsl #2
\r
1803 orr r8, r8, r10,lsl #15
\r
1804 add r8, r8, r12,lsl #15 @ pix_d 6, 7
\r
1805 sub r10,r10,r10,lsr #2 @ r10= 3/4 pix_s 2
\r
1806 and r7, r7, r9, lsl #2
\r
1807 add r10,r10,r7, lsr #2 @ += 1/4 pix_s 3
\r
1808 orr r10,r10,r7, lsl #16 @ pix_d 8, 9
\r
1812 stmia r0!, {r4,r5,r6,r8,r10}
\r
1813 bne .fl_loop32scale_RGB555
\r
1815 ldmfd sp!, {r9,r10}
\r
1816 ldmfd sp!, {r4-r8,lr}
\r
1819 .if UNALIGNED_DRAWLINEDEST
\r
1820 @ unaligned versions of loops
\r
1821 @ warning: starts drawing 2bytes before dst
\r
1824 sub r0, r0, #2 @ initial adjustment
\r
1831 and r6, lr, r12,lsl #1
\r
1833 and r5, lr, r12,lsr #7
\r
1835 orr r4, r8, r6, lsl #16
\r
1837 and r6, lr, r12,lsr #15
\r
1839 and r8, lr, r12,lsr #23
\r
1841 orr r5, r5, r6, lsl #16
\r
1843 and r6, lr, r7, lsl #1
\r
1845 and r12,lr, r7, lsr #7
\r
1846 ldrh r12,[r3, r12]
\r
1847 orr r6, r8, r6, lsl #16
\r
1849 and r8, lr, r7, lsr #15
\r
1851 and r7, lr, r7, lsr #23
\r
1854 orr r12,r12,r8, lsl #16
\r
1857 stmia r0!, {r4,r5,r6,r12}
\r
1858 bne .fl_loopRGB555u
\r
1862 ldmfd sp!, {r4-r8,lr}
\r
1866 .fl_32scale_RGB555u:
\r
1867 sub r0, r0, #2 @ initial adjustment
\r
1870 @ r9 f800 07e0 001f | e000 0780 001c | 3800 01e0 0007
\r
1871 .fl_loop32scale_RGB555u:
\r
1875 and r6, lr, r12,lsl #1
\r
1877 and r5, lr, r12,lsr #7
\r
1879 and r6, r6, r9, lsl #2
\r
1880 orr r4, r4, r6, lsl #16 @ r4 = pix_d -1, 0
\r
1882 and r5, r5, r9, lsl #2
\r
1883 sub r8, r5, r5, lsr #2 @ r8 = 3/4 pix_s 1
\r
1884 add r6, r8, r6, lsr #2 @ r6 = (1/4 pix_s 0) + (3/4 pix_s 1)
\r
1885 orr r5, r6, r5, lsl #15
\r
1887 and r6, lr, r12,lsr #15
\r
1889 and r12,lr, r12,lsr #23
\r
1890 ldrh r12,[r3, r12]
\r
1891 and r6, r6, r9, lsl #2
\r
1892 add r5, r5, r6, lsl #15 @ r5 = pix_d 1, 2
\r
1894 and r8, lr, r7, lsl #1
\r
1896 and r10,lr, r7, lsr #7
\r
1897 ldrh r10,[r3, r10]
\r
1898 and r12,r12,r9, lsl #2
\r
1899 sub r6, r6, r6, lsr #2 @ r6 = 3/4 pix_s 2
\r
1900 add r6, r6, r12,lsr #2
\r
1901 orr r6, r6, r12,lsl #16 @ r6 = pix_d 3, 4
\r
1903 and r8, r8, r9, lsl #2
\r
1904 and r10,r10,r9, lsl #2
\r
1905 sub r12,r10,r10,lsr #2 @ r12 = 3/4 pix_s 5
\r
1906 orr r8, r8, r8, lsl #14
\r
1907 add r8, r8, r12,lsl #16 @ r8 = pix_d 5, 6
\r
1908 and r12,lr, r7, lsr #15
\r
1909 ldrh r12,[r3, r12]
\r
1910 and r7, lr, r7, lsr #23
\r
1912 and r12,r12,r9, lsl #2
\r
1914 mov r10,r10, lsr #1
\r
1915 sub r12,r12,r12,lsr #2 @ r12 = 3/4 pix_s 6
\r
1916 orr r10,r10,r12,lsl #16
\r
1917 and r7, r7, r9, lsl #2
\r
1918 add r10,r10,r7, lsl #14 @ r10 = pix_d 7, 8
\r
1922 stmia r0!, {r4,r5,r6,r8,r10}
\r
1924 bne .fl_loop32scale_RGB555u
\r
1928 ldmfd sp!, {r9,r10}
\r
1929 ldmfd sp!, {r4-r8,lr}
\r
1932 .endif @ UNALIGNED_DRAWLINEDEST
\r
1935 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1938 .global blockcpy @ void *dst, void *src, size_t n
\r
1941 stmfd sp!, {r4,r5}
\r
1942 mov r2, r2, lsr #4
\r
1944 ldmia r1!, {r3-r5,r12}
\r
1946 stmia r0!, {r3-r5,r12}
\r
1948 ldmfd sp!, {r4,r5}
\r
1952 .global blockcpy_or @ void *dst, void *src, size_t n, int pat
\r
1955 stmfd sp!, {r4-r6}
\r
1956 orr r3, r3, r3, lsl #8
\r
1957 orr r3, r3, r3, lsl #16
\r
1958 mov r2, r2, lsr #4
\r
1960 ldmia r1!, {r4-r6,r12}
\r
1966 stmia r0!, {r4-r6,r12}
\r
1967 bne blockcpy_loop_or
\r
1968 ldmfd sp!, {r4-r6}
\r