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
18 .extern DrawLineDest
\r
19 .extern DrawStripInterlace
\r
20 .extern HighCacheS_ptr
\r
22 .equ PDRAW_SPRITES_MOVED, (1<<0)
\r
23 .equ PDRAW_WND_DIFF_PRIO, (1<<1)
\r
24 .equ PDRAW_ACC_SPRITES, (1<<2)
\r
25 .equ PDRAW_DIRTY_SPRITES, (1<<4)
\r
26 .equ PDRAW_PLANE_HI_PRIO, (1<<6)
\r
27 .equ PDRAW_SHHI_DONE, (1<<7)
\r
28 .equ MAX_LINE_SPRITES, 30
\r
31 .macro TilePixel pat lsrr offs
\r
35 ands r4, \pat, r2, lsr #\lsrr
\r
38 strneb r4, [r1,#\offs]
\r
41 @ TileNorm (r1=pdest, r2=pixels8, r3=pal) r4: scratch, pat: register with helper pattern 0xf
\r
43 TilePixel \pat, 12, 0 @ #0x0000f000
\r
44 TilePixel \pat, 8, 1 @ #0x00000f00
\r
45 TilePixel \pat, 4, 2 @ #0x000000f0
\r
46 TilePixel \pat, 0, 3 @ #0x0000000f
\r
47 TilePixel \pat, 28, 4 @ #0xf0000000
\r
48 TilePixel \pat, 24, 5 @ #0x0f000000
\r
49 TilePixel \pat, 20, 6 @ #0x00f00000
\r
50 TilePixel \pat, 16, 7 @ #0x000f0000
\r
53 @ TileFlip (r1=pdest, r2=pixels8, r3=pal) r4: scratch, pat: register with helper pattern 0xf
\r
55 TilePixel \pat, 16, 0 @ #0x000f0000
\r
56 TilePixel \pat, 20, 1 @ #0x00f00000
\r
57 TilePixel \pat, 24, 2 @ #0x0f000000
\r
58 TilePixel \pat, 28, 3 @ #0xf0000000
\r
59 TilePixel \pat, 0, 4 @ #0x0000000f
\r
60 TilePixel \pat, 4, 5 @ #0x000000f0
\r
61 TilePixel \pat, 8, 6 @ #0x00000f00
\r
62 TilePixel \pat, 12, 7 @ #0x0000f000
\r
65 @ shadow/hilight mode
\r
67 @ this one is for hi priority layer
\r
68 .macro TilePixelShHP lsrr offs
\r
72 ands r4, r12, r2, lsr #\lsrr
\r
74 ldreqb r4, [r1,#\offs]
\r
77 strb r4, [r1,#\offs]
\r
80 @ TileNormShHP (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: register with helper pattern 0xf, touches r3 high bits
\r
82 TilePixelShHP 12, 0 @ #0x0000f000
\r
83 TilePixelShHP 8, 1 @ #0x00000f00
\r
84 TilePixelShHP 4, 2 @ #0x000000f0
\r
85 TilePixelShHP 0, 3 @ #0x0000000f
\r
86 TilePixelShHP 28, 4 @ #0xf0000000
\r
87 TilePixelShHP 24, 5 @ #0x0f000000
\r
88 TilePixelShHP 20, 6 @ #0x00f00000
\r
89 TilePixelShHP 16, 7 @ #0x000f0000
\r
92 @ TileFlipShHP (r1=pdest, r2=pixels8, r3=pal) r4: scratch, pat: register with helper pattern 0xf
\r
94 TilePixelShHP 16, 0 @ #0x000f0000
\r
95 TilePixelShHP 20, 1 @ #0x00f00000
\r
96 TilePixelShHP 24, 2 @ #0x0f000000
\r
97 TilePixelShHP 28, 3 @ #0xf0000000
\r
98 TilePixelShHP 0, 4 @ #0x0000000f
\r
99 TilePixelShHP 4, 5 @ #0x000000f0
\r
100 TilePixelShHP 8, 6 @ #0x00000f00
\r
101 TilePixelShHP 12, 7 @ #0x0000f000
\r
105 @ TileSingleSh (r1=pdest, r2=pixels8, r3=pal) r4,r7: scratch, r0=sx; r12: helper pattern 0xf
\r
106 .macro TileSingleSh
\r
107 tst r0, #1 @ not aligned?
\r
113 strneb r4, [r1], #1
\r
114 streqh r4, [r1], #2
\r
126 strneb r4, [r1], #1
\r
129 @ TileSingleHi (r1=pdest, r2=pixels8, r3=pal) r4,r7: scratch, r0=sx, r12: register with helper pattern 0xf
\r
130 .macro TileSingleHi
\r
131 tst r1, #1 @ not aligned?
\r
134 ldrneb r4, [r1], #1
\r
135 ldreqh r4, [r1], #2 @ 1ci
\r
137 bic r4, r4, r7, lsr #1
\r
139 strneb r4, [r1, #-3]
\r
140 streqh r4, [r1, #-4]
\r
142 bic r12, r12, r7, lsr #1
\r
144 strh r12, [r1, #-4]
\r
146 bic r4, r4, r7, lsr #1
\r
150 bic r12, r12, r7, lsr #1
\r
152 strh r12, [r1, #-2]
\r
153 bicne r4, r4, r7, lsr #1
\r
155 strneb r4, [r1], #1
\r
159 .macro TileDoShGenPixel shift ofs
\r
161 ands r4, r12, r2, lsr #\shift
\r
167 ldrgeb r4, [r1,#\ofs]
\r
168 orrlt r4, r3, r4 @ normal
\r
170 biceq r4, r4, #0xc0 @ hilight
\r
171 orreq r4, r4, #0x80
\r
172 orrgt r4, r4, #0xc0 @ shadow
\r
174 strb r4, [r1,#\ofs]
\r
178 @ TileFlipSh (r1=pdest, r2=pixels8, r3=pal) r4,r7: scratch, r0=sx, r12: register with helper pattern 0xf
\r
180 TileDoShGenPixel 16, 0 @ #0x000f0000
\r
181 TileDoShGenPixel 20, 1 @ #0x00f00000
\r
182 TileDoShGenPixel 24, 2 @ #0x0f000000
\r
183 TileDoShGenPixel 28, 3 @ #0xf0000000
\r
184 TileDoShGenPixel 0, 4 @ #0x0000000f
\r
185 TileDoShGenPixel 4, 5 @ #0x000000f0
\r
186 TileDoShGenPixel 8, 6 @ #0x00000f00
\r
187 TileDoShGenPixel 12, 7 @ #0x0000f000
\r
190 @ TileNormSh (r1=pdest, r2=pixels8, r3=pal) r4,r7: scratch, r0=sx, r12: register with helper pattern 0xf
\r
192 TileDoShGenPixel 12, 0 @ #0x0000f000
\r
193 TileDoShGenPixel 8, 1 @ #0x00000f00
\r
194 TileDoShGenPixel 4, 2 @ #0x000000f0
\r
195 TileDoShGenPixel 0, 3 @ #0x0000000f
\r
196 TileDoShGenPixel 28, 4 @ #0xf0000000
\r
197 TileDoShGenPixel 24, 5 @ #0x0f000000
\r
198 TileDoShGenPixel 20, 6 @ #0x00f00000
\r
199 TileDoShGenPixel 16, 7 @ #0x000f0000
\r
202 .macro TileDoShGenPixel_noop shift ofs
\r
204 and r4, r12, r2, lsr #\shift
\r
210 orrcc r4, r3, r4 @ 0-0xc (was 1-0xd)
\r
211 strccb r4, [r1,#\ofs]
\r
214 .macro TileFlipSh_noop
\r
215 TileDoShGenPixel_noop 16, 0 @ #0x000f0000
\r
216 TileDoShGenPixel_noop 20, 1 @ #0x00f00000
\r
217 TileDoShGenPixel_noop 24, 2 @ #0x0f000000
\r
218 TileDoShGenPixel_noop 28, 3 @ #0xf0000000
\r
219 TileDoShGenPixel_noop 0, 4 @ #0x0000000f
\r
220 TileDoShGenPixel_noop 4, 5 @ #0x000000f0
\r
221 TileDoShGenPixel_noop 8, 6 @ #0x00000f00
\r
222 TileDoShGenPixel_noop 12, 7 @ #0x0000f000
\r
225 .macro TileNormSh_noop
\r
226 TileDoShGenPixel_noop 12, 0 @ #0x0000f000
\r
227 TileDoShGenPixel_noop 8, 1 @ #0x00000f00
\r
228 TileDoShGenPixel_noop 4, 2 @ #0x000000f0
\r
229 TileDoShGenPixel_noop 0, 3 @ #0x0000000f
\r
230 TileDoShGenPixel_noop 28, 4 @ #0xf0000000
\r
231 TileDoShGenPixel_noop 24, 5 @ #0x0f000000
\r
232 TileDoShGenPixel_noop 20, 6 @ #0x00f00000
\r
233 TileDoShGenPixel_noop 16, 7 @ #0x000f0000
\r
236 .macro TileDoShGenPixel_onlyop_lp shift ofs
\r
238 ands r7, r12, r2, lsr #\shift
\r
242 ldrneb r4, [r1,#\ofs]
\r
247 biceq r4, r4, #0xc0 @ hilight
\r
248 orreq r4, r4, #0x80
\r
249 orrgt r4, r4, #0xc0 @ shadow
\r
250 strgeb r4, [r1,#\ofs]
\r
254 .macro TileFlipSh_onlyop_lp
\r
255 TileDoShGenPixel_onlyop_lp 16, 0 @ #0x000f0000
\r
256 TileDoShGenPixel_onlyop_lp 20, 1 @ #0x00f00000
\r
257 TileDoShGenPixel_onlyop_lp 24, 2 @ #0x0f000000
\r
258 TileDoShGenPixel_onlyop_lp 28, 3 @ #0xf0000000
\r
259 TileDoShGenPixel_onlyop_lp 0, 4 @ #0x0000000f
\r
260 TileDoShGenPixel_onlyop_lp 4, 5 @ #0x000000f0
\r
261 TileDoShGenPixel_onlyop_lp 8, 6 @ #0x00000f00
\r
262 TileDoShGenPixel_onlyop_lp 12, 7 @ #0x0000f000
\r
265 .macro TileNormSh_onlyop_lp
\r
266 TileDoShGenPixel_onlyop_lp 12, 0 @ #0x0000f000
\r
267 TileDoShGenPixel_onlyop_lp 8, 1 @ #0x00000f00
\r
268 TileDoShGenPixel_onlyop_lp 4, 2 @ #0x000000f0
\r
269 TileDoShGenPixel_onlyop_lp 0, 3 @ #0x0000000f
\r
270 TileDoShGenPixel_onlyop_lp 28, 4 @ #0xf0000000
\r
271 TileDoShGenPixel_onlyop_lp 24, 5 @ #0x0f000000
\r
272 TileDoShGenPixel_onlyop_lp 20, 6 @ #0x00f00000
\r
273 TileDoShGenPixel_onlyop_lp 16, 7 @ #0x000f0000
\r
277 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
281 @ int nametab; // 0x00
\r
282 @ int line; // 0x04
\r
283 @ int hscroll; // 0x08
\r
284 @ int xmask; // 0x0C
\r
285 @ int *hc; // 0x10 (pointer to cache buffer)
\r
286 @ int cells; // 0x14
\r
289 @ void DrawLayer(int plane_sh, int *hcache, int cellskip, int maxcells);
\r
294 stmfd sp!, {r4-r11,lr}
\r
296 ldr r11, =(Pico+0x22228) @ Pico.video
\r
299 ldrb r7, [r11, #16] @ ??vv??hh
\r
301 mov r6, r1 @ hcache
\r
302 orr r9, r3, r0, lsl #30
\r
303 orr r9, r9, r2, lsl #8 @ r9=sh[31]|cellskip[15:8]|maxcells[7:0] (tmp)
\r
306 orr r1, r1, #0x00ff
\r
310 biclt r1, r1, #0xfc00
\r
311 biceq r1, r1, #0xfe00
\r
312 bicgt r1, r1, #0xff00 @ r1=ymask=(height<<8)|0xff; ...; // Y Mask in pixels
\r
316 subge r10, r10, #1 @ r10=shift[width] (5,6,6,7)
\r
319 mov r5, r8, lsl r10
\r
320 sub r5, r5, #1 @ r5=xmask
\r
324 ldreqb r12, [r11, #2]
\r
325 ldrneb r12, [r11, #4]
\r
327 ldr r2, =DrawScanline @ trying to make good use of pipeline here
\r
328 ldr lr, =(Pico+0x10000) @ lr=Pico.vram
\r
330 moveq r12, r12, lsl #10
\r
331 movne r12, r12, lsl #13
\r
332 and r12, r12, #(7<<13) @ r12=(ts->nametab<<1) (halfword compliant)
\r
334 ldrh r8, [r11, #12]
\r
335 ldrb r7, [r11, #11]
\r
338 mov r4, r8, lsr #8 @ pvid->reg[13]
\r
339 mov r4, r4, lsl #10 @ htab=pvid->reg[13]<<9; (halfwords)
\r
341 addne r4, r4, r2, lsl #2 @ htab+=DrawScanline<<1; // Offset by line
\r
343 biceq r4, r4, #0x1f @ htab&=~0xf; // Offset by tile
\r
344 add r4, r4, r0, lsl #1 @ htab+=plane
\r
345 bic r4, r4, #0x00ff0000 @ just in case
\r
346 ldrh r3, [lr, r4] @ r3=hscroll
\r
349 bne .DrawStrip_vsscroll
\r
351 @ Get vertical scroll value:
\r
352 add r7, lr, #0x012000
\r
353 add r7, r7, #0x000180 @ r7=Pico.vsram (Pico+0x22180)
\r
358 bne .DrawStrip_interlace
\r
361 movne r7, r7, lsr #16
\r
363 @ Find the line in the name table
\r
367 add r10, r10, #1 @ shift[width]++
\r
368 add r12, r12, r4, lsl r10 @ nametab+=(ts.line>>3)<<shift[width];
\r
370 @ ldmia r0, {r1,r2,r3,r5,r6,r9} @ r2=line, r3=ts->hscroll, r5=ts->xmask, r6=ts->hc, r9=ts->cells
\r
373 mov r10,r10, lsl #1 @ r10=ty=(ts->line&7)<<1;
\r
374 orr r10,r10, r9, lsl #24
\r
377 mov r8, r8, lsr #3 @ r8=tilex=(-ts->hscroll)>>3
\r
381 add r7, r1, #1 @ r7=dx=((ts->hscroll-1)&7)+1
\r
385 orrne r10,r10, #1<<23 @ r10=(cells<<24|sh<<23|hi_not_empty<<22|had_output<<21|ty)
\r
386 movne r3, #0x40 @ default to shadowed pal on sh mode
\r
389 addne r10,r10, #0x01000000 @ we will loop cells+1 times if there is scroll
\r
391 and r9, r9, #0xff00
\r
392 add r8, r8, r9, lsr #8 @ tilex+=cellskip
\r
393 add r7, r7, r9, lsr #5 @ dx+=cellskip<<3;
\r
394 sub r10,r10,r9, lsl #16 @ cells-=cellskip
\r
396 @ cache some stuff to avoid mem access
\r
397 .if OVERRIDE_HIGHCOL
\r
406 mvn r9, #0 @ r9=prevcode=-1
\r
407 add r1, r11, r7 @ r1=pdest
\r
410 @ r4 & r7 are scratch in this loop
\r
413 .dsloop: @ 40-41 times
\r
414 subs r10,r10, #0x01000000
\r
419 add r7, lr, r7, lsl #1 @ Pico.vram+((tilex&ts->xmask) as halfwords)
\r
420 ldrh r7, [r7, r12] @ r7=code (int, but from unsigned, no sign extend)
\r
426 bne .DrawStrip_hiprio
\r
429 beq .DrawStrip_samecode @ we know stuff about this tile already
\r
431 mov r9, r7 @ remember code
\r
432 orr r10, r10, #1<<21 @ seen non hi-prio tile
\r
434 movs r2, r9, lsl #20 @ if (code&0x1000)
\r
436 add r2, r2, r10, lsl #17
\r
437 mov r2, r2, lsr #17
\r
438 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
440 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
443 and r3, r9, #0x6000
\r
444 add r3, r7, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);
\r
446 .DrawStrip_samecode:
\r
448 beq .dsloop @ tileline blank
\r
451 beq .DrawStrip_SingleColor @ tileline singlecolor
\r
454 bne .DrawStrip_TileFlip
\r
456 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r0: helper pattern
\r
457 .DrawStrip_TileNorm:
\r
461 .DrawStrip_TileFlip:
\r
465 .DrawStrip_SingleColor:
\r
468 orr r4, r4, r4, lsl #8
\r
469 tst r1, #1 @ not aligned?
\r
470 strneb r4, [r1], #1
\r
471 streqh r4, [r1], #2
\r
475 strneb r4, [r1], #1 @ have a remaining unaligned pixel?
\r
478 .DrawStrip_hiprio_maybempt:
\r
480 beq .dsloop @ must've been empty, otherwise we wouldn't get here
\r
481 movs r2, r7, lsl #20 @ if (code&0x1000)
\r
483 add r2, r2, r10, lsl #17
\r
484 mov r2, r2, lsr #17
\r
485 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
486 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
487 mov r9, r7 @ remember code
\r
490 orr r10, r10, #1<<22
\r
493 tst r10, #0x00c00000
\r
494 beq .DrawStrip_hiprio_maybempt
\r
496 orr r7, r7, r0, lsl #16
\r
497 orr r7, r7, r10, lsl #25 @ (ty<<25)
\r
499 eorne r7, r7, #7<<26 @ if(code&0x1000) cval^=7<<26;
\r
500 str r7, [r6], #4 @ cache hi priority tile
\r
505 tst r10, #1<<21 @ seen non hi-prio tile
\r
506 ldreq r1, =rendstatus
\r
509 str r0, [r6] @ terminate the cache list
\r
510 orreq r2, r2, #PDRAW_PLANE_HI_PRIO @ had a layer with all hi-prio tiles
\r
513 ldmfd sp!, {r4-r11,lr}
\r
516 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
518 .DrawStrip_vsscroll:
\r
520 mov r8, r8, lsr #3 @ r8=tilex=(-ts->hscroll)>>3
\r
521 bic r8, r8, #0x3fc00000
\r
522 orr r8, r8, r5, lsl #25 @ r8=(xmask[31:25]|had_output[24]|tilex[21:0])
\r
524 ldr r4, =DrawScanline
\r
525 orr r5, r1, r10, lsl #24
\r
528 orr r5, r5, r4, lsl #16 @ r5=(shift_width[31:24]|scanline[23:16]|ymask[15:0])
\r
530 add r7, r1, #1 @ r7=dx=((ts->hscroll-1)&7)+1
\r
532 mov r10,r9, lsl #16
\r
534 orrne r10,r10, #0x8000
\r
537 orr r10,r10, #0xff000000 @ will be adjusted on entering loop
\r
538 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
539 movne r3, #0x40 @ default to shadowed pal on sh mode
\r
542 subne r10,r10, #0x01000000 @ have hscroll, start with negative cell
\r
544 and r9, r9, #0xff00
\r
545 add r8, r8, r9, lsr #8 @ tilex+=cellskip
\r
546 add r7, r7, r9, lsr #5 @ dx+=cellskip<<3;
\r
547 add r10,r10,r9, lsl #16 @ cell+=cellskip
\r
549 @ cache some stuff to avoid mem access
\r
550 .if OVERRIDE_HIGHCOL
\r
559 mvn r9, #0 @ r9=prevcode=-1
\r
560 add r1, r11, r7 @ r1=pdest
\r
562 @ r4 & r7 are scratch in this loop
\r
565 .dsloop_vs: @ 40-41 times
\r
566 add r10,r10, #0x01000000
\r
567 and r4, r10, #0x003f0000
\r
568 cmp r4, r10, asr #8
\r
569 ble .dsloop_vs_exit
\r
571 @ calc offset and read tileline code to r7, also calc ty
\r
572 add r7, lr, #0x012000
\r
573 add r7, r7, #0x000180 @ r7=Pico.vsram (Pico+0x22180)
\r
574 add r7, r7, r10,asr #23 @ vsram + ((cell&~1)<<1)
\r
576 tst r10,#0x8000 @ plane1?
\r
578 ldrh r7, [r7] @ r7=vscroll
\r
580 bic r10,r10,#0xff @ clear old ty
\r
581 and r4, r5, #0xff0000 @ scanline
\r
582 add r4, r4, r7, lsl #16 @ ... += vscroll
\r
583 and r4, r4, r5, lsl #16 @ ... &= ymask
\r
584 and r7, r4, #0x70000
\r
585 orr r10,r10,r7, lsr #15 @ new ty
\r
587 mov r4, r4, lsr #19
\r
588 mov r7, r5, lsr #24
\r
589 mov r4, r4, lsl r7 @ nametabadd
\r
591 and r7, r8, r8, lsr #25
\r
592 add r7, lr, r7, lsl #1 @ Pico.vram+((tilex&ts->xmask) as halfwords)
\r
593 add r7, r7, r4, lsl #1
\r
594 ldrh r7, [r7, r12] @ r7=code (int, but from unsigned, no sign extend)
\r
600 bne .DrawStrip_vs_hiprio
\r
603 beq .DrawStrip_vs_samecode @ we know stuff about this tile already
\r
605 mov r9, r7 @ remember code
\r
606 orr r8, r8, #(1<<24)@ seen non hi-prio tile
\r
608 movs r2, r9, lsl #20 @ if (code&0x1000)
\r
610 add r2, r2, r10, lsl #17
\r
611 mov r2, r2, lsr #17
\r
612 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
614 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
617 and r3, r9, #0x6000
\r
618 add r3, r7, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);
\r
620 .DrawStrip_vs_samecode:
\r
622 beq .dsloop_vs @ tileline blank
\r
625 beq .DrawStrip_vs_SingleColor @ tileline singlecolor
\r
628 bne .DrawStrip_vs_TileFlip
\r
630 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r0: helper pattern
\r
631 .DrawStrip_vs_TileNorm:
\r
635 .DrawStrip_vs_TileFlip:
\r
639 .DrawStrip_vs_SingleColor:
\r
642 orr r4, r4, r4, lsl #8
\r
643 tst r1, #1 @ not aligned?
\r
644 strneb r4, [r1], #1
\r
645 streqh r4, [r1], #2
\r
649 strneb r4, [r1], #1 @ have a remaining unaligned pixel?
\r
652 .DrawStrip_vs_hiprio:
\r
653 tst r10, #0x00c00000
\r
654 beq .DrawStrip_vs_hiprio_maybempt
\r
656 orr r7, r7, r0, lsl #16
\r
657 orr r7, r7, r10, lsl #25 @ (ty<<25)
\r
659 eorne r7, r7, #7<<26 @ if(code&0x1000) cval^=7<<26;
\r
660 str r7, [r6], #4 @ cache hi priority tile
\r
664 .DrawStrip_vs_hiprio_maybempt:
\r
666 beq .dsloop_vs @ must've been empty, otherwise we wouldn't get here
\r
667 movs r2, r7, lsl #20 @ if (code&0x1000)
\r
669 add r2, r2, r10, lsl #17
\r
670 mov r2, r2, lsr #17
\r
671 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
672 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
673 mov r9, r7 @ remember code
\r
675 orrne r10, r10, #1<<22
\r
676 bne .DrawStrip_vs_hiprio
\r
680 tst r8, #(1<<24) @ seen non hi-prio tile
\r
681 ldreq r1, =rendstatus
\r
684 str r0, [r6] @ terminate the cache list
\r
685 orreq r2, r2, #PDRAW_PLANE_HI_PRIO @ had a layer with all hi-prio tiles
\r
688 ldmfd sp!, {r4-r11,lr}
\r
692 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
694 @ interlace mode 2? Sonic 2?
\r
695 .DrawStrip_interlace:
\r
697 moveq r7, r7, lsl #21
\r
698 movne r7, r7, lsl #5
\r
700 @ Find the line in the name table
\r
701 add r2, r7, r2, lsl #22 @ r2=(vscroll+(DrawScanline<<1))<<21 (11 bits);
\r
702 orr r1, r1, #0x80000000
\r
703 and r2, r2, r1, ror #10 @ &((ymask<<1)|1)<<21;
\r
704 mov r2, r2, lsr #21
\r
706 mov r12, r12, lsr #1 @ halfwords
\r
707 add r0, r12, r4, lsl r10 @ nametab+=(ts.line>>4)<<shift[width];
\r
711 stmia sp, {r0,r2,r3,r5,r6,r9}
\r
714 bl DrawStripInterlace @ struct TileStrip *ts
\r
717 ldmfd sp!, {r4-r11,lr}
\r
722 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
725 .global BackFill @ int reg7, int sh
\r
728 stmfd sp!, {r4-r9,lr}
\r
730 .if OVERRIDE_HIGHCOL
\r
732 mov r0, r0, lsl #26
\r
734 mov r0, r0, lsr #26
\r
737 ldr lr, =(HighCol+8)
\r
738 mov r0, r0, lsl #26
\r
739 mov r0, r0, lsr #26
\r
742 orr r0, r0, r1, lsl #6
\r
743 orr r0, r0, r0, lsl #8
\r
744 orr r0, r0, r0, lsl #16
\r
755 stmia lr!, {r0-r7} @ 10*8*4
\r
766 ldmfd sp!, {r4-r9,r12}
\r
770 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
773 .global DrawTilesFromCache @ int *hc, int sh, int rlim
\r
775 DrawTilesFromCache:
\r
776 stmfd sp!, {r4-r8,r11,lr}
\r
778 @ cache some stuff to avoid mem access
\r
779 .if OVERRIDE_HIGHCOL
\r
787 ldr lr, =(Pico+0x10000) @ lr=Pico.vram
\r
789 mvn r5, #0 @ r5=prevcode=-1
\r
791 orr r8, r8, r2, lsl #1
\r
792 bne .dtfc_check_rendflags
\r
796 ldr r6, [r0], #4 @ read code
\r
797 movs r1, r6, lsr #16 @ r1=dx;
\r
798 ldmeqfd sp!, {r4-r8,r11,pc} @ dx is never zero, this must be a terminator, return
\r
799 bic r4, r1, #0xfe00
\r
800 add r1, r11, r4 @ r1=pdest
\r
802 mov r7, r6, lsl #16
\r
803 cmp r5, r7, lsr #16
\r
804 beq .dtfc_samecode @ if (code==prevcode)
\r
806 mov r5, r7, lsr #16
\r
808 mov r2, r5, lsl #21
\r
809 mov r2, r2, lsr #17 @ r2=addr=(code&0x7ff)<<4;
\r
810 add r2, r2, r6, lsr #25 @ addr+=ty
\r
812 and r3, r5, #0x6000
\r
813 mov r3, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);
\r
815 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
818 rsbs r4, r4, r8, lsr #1
\r
828 beq .dtfc_SingleColor @ tileline singlecolor
\r
833 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
845 orr r4, r4, r4, lsl #8
\r
846 tst r1, #1 @ not aligned?
\r
847 strneb r4, [r1], #1
\r
848 streqh r4, [r1], #2
\r
852 strneb r4, [r1], #1 @ have a remaining unaligned pixel?
\r
857 beq .dtfc_shadow_blank
\r
860 beq .dtfc_SingleColor @ tileline singlecolor
\r
863 bne .dtfc_TileFlipShHP
\r
865 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
866 .dtfc_TileNormShHP:
\r
870 .dtfc_TileFlipShHP:
\r
874 .dtfc_shadow_blank:
\r
879 strneb r4, [r1], #1
\r
881 orr r6, r6, r6, lsl #8
\r
897 add r4, r4, #7 @ 0-6
\r
901 mov r2, r2, ror #16
\r
902 tst r5, #0x0800 @ flipped?
\r
905 mov r2, r2, ror #16
\r
915 @ check if we have detected layer covered with hi-prio tiles:
\r
916 .dtfc_check_rendflags:
\r
917 ldr r1, =rendstatus
\r
919 tst r2, #(PDRAW_PLANE_HI_PRIO|PDRAW_SHHI_DONE)
\r
921 bic r8, r8, #1 @ sh/hi mode off
\r
922 tst r2, #PDRAW_SHHI_DONE
\r
923 bne .dtfc_loop @ already processed
\r
924 orr r2, r2, #PDRAW_SHHI_DONE
\r
930 orr r6, r6, r6, lsl #8
\r
931 orr r6, r6, r6, lsl #16
\r
933 ldmia r1, {r2,r4,r5,r7}
\r
939 stmia r1!,{r2,r4,r5,r7}
\r
940 bne .dtfc_loop_shprep
\r
942 mvn r5, #0 @ r5=prevcode=-1
\r
947 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
950 .global DrawSpritesSHi @ int prio_unused, int sh
\r
953 ldr r3, =DrawScanline
\r
956 mov r3, #(MAX_LINE_SPRITES+2)
\r
957 mla r2, r12, r3, r2
\r
962 stmfd sp!, {r4-r11,lr}
\r
964 strb r12,[r2,#1] @ set end marker
\r
966 add r10,r10,r3 @ r10=HighLnSpr end
\r
968 str r1, [sp, #-4] @ no calls after this point
\r
970 .if OVERRIDE_HIGHCOL
\r
978 ldr lr, =(Pico+0x10000) @ lr=Pico.vram
\r
983 ldrb r0, [r10,#-1]!
\r
984 ldr r1, =HighPreSpr
\r
987 ldmeqfd sp!, {r4-r11,pc} @ end of list
\r
989 add r0, r1, r0, lsl #3
\r
991 ldr r9, [r0, #4] @ sprite[1]
\r
992 mov r2, r9, asr #16 @ r2=sx
\r
994 mov r9, r9, lsl #16
\r
995 mov r3, r9, lsr #31 @ priority
\r
996 mov r9, r9, lsr #16
\r
997 orr r9, r9, r8, lsl #31 @ r9=code|sh[31]
\r
998 and r4, r9, #0x6000
\r
999 orr r9, r9, r4, lsl #16
\r
1000 orr r9, r9, #0x10000000 @ r9=scc1 ???? ... <code> (s=shadow/hilight, cc=pal)
\r
1001 cmp r12,r9, lsr #28 @ sh/hi with pal3?
\r
1002 cmpne r3, #1 @ if not, is hi prio
\r
1003 bne DrawSpriteSHi @ non-operator low sprite, already drawn
\r
1005 ldr r3, [r0] @ sprite[0]
\r
1006 ldr r7, =DrawScanline
\r
1007 mov r6, r3, lsr #28
\r
1008 sub r6, r6, #1 @ r6=width-1 (inc later)
\r
1009 mov r5, r3, lsr #24
\r
1010 and r5, r5, #7 @ r5=height
\r
1012 mov r0, r3, lsl #16 @ r4=sy<<16 (tmp)
\r
1015 sub r7, r7, r0, asr #16 @ r7=row=DrawScanline-sy
\r
1018 movne r0, r5, lsl #3
\r
1020 subne r7, r0, r7 @ if (code&0x1000) row=(height<<3)-1-row; // Flip Y
\r
1022 add r8, r9, r7, lsr #3 @ tile+=row>>3; // Tile number increases going down
\r
1024 mlane r8, r5, r6, r8 @ if (code&0x0800) { tile+=delta*(width-1);
\r
1025 rsbne r5, r5, #0 @ delta=-delta; } // r5=delta now
\r
1027 mov r8, r8, lsl #21
\r
1028 mov r8, r8, lsr #17
\r
1030 add r8, r8, r7, lsl #1 @ tile+=(row&7)<<1; // Tile address
\r
1032 mov r5, r5, lsl #4 @ delta<<=4; // Delta of address
\r
1034 orrs r3, r9, r9, lsl #4
\r
1035 mov r3, r4, lsr #9 @ r3=pal=((code>>9)&0x30);
\r
1037 add r6, r6, #1 @ inc now
\r
1038 adds r0, r2, #0 @ mov sx to r0 and set ZV flags
\r
1039 b .dsprShi_loop_enter
\r
1042 subs r6, r6, #1 @ width--
\r
1044 adds r0, r0, #8 @ sx+=8
\r
1045 add r8, r8, r5 @ tile+=delta
\r
1047 .dsprShi_loop_enter:
\r
1048 ble .dsprShi_loop @ sx <= 0
\r
1052 mov r8, r8, lsl #17
\r
1053 mov r8, r8, lsr #17 @ tile&=0x7fff; // Clip tile address
\r
1055 ldr r2, [lr, r8, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
1056 add r1, r11, r0 @ r1=pdest
\r
1060 cmp r12, r9, lsr #28
\r
1061 beq .dsprShi_shadow
\r
1063 cmp r2, r2, ror #4
\r
1064 beq .dsprShi_SingleColor @ tileline singlecolor
\r
1067 bne .dsprShi_TileFlip
\r
1069 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
1071 .dsprShi_TileNorm:
\r
1075 .dsprShi_TileFlip:
\r
1079 .dsprShi_SingleColor:
\r
1082 orr r4, r4, r4, lsl #8
\r
1083 tst r0, #1 @ not aligned?
\r
1084 strneb r4, [r1], #1
\r
1085 streqh r4, [r1], #2
\r
1089 strneb r4, [r1], #1
\r
1094 beq .dsprShi_shadow_lowpri
\r
1096 cmp r2, r2, ror #4
\r
1097 beq .dsprShi_singlec_sh
\r
1100 bne .dsprShi_TileFlip_sh
\r
1102 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
1103 .dsprShi_TileNorm_sh:
\r
1107 .dsprShi_TileFlip_sh:
\r
1111 .dsprShi_singlec_sh:
\r
1112 cmp r2, #0xe0000000
\r
1113 bcc .dsprShi_SingleColor @ normal singlecolor tileline (carry inverted in ARM)
\r
1114 tst r2, #0x10000000
\r
1115 bne .dsprShi_sh_sh
\r
1123 .dsprShi_shadow_lowpri:
\r
1125 bne .dsprShi_TileFlip_sh_lp
\r
1127 .dsprShi_TileNorm_sh_lp:
\r
1128 TileNormSh_onlyop_lp
\r
1131 .dsprShi_TileFlip_sh_lp:
\r
1132 TileFlipSh_onlyop_lp
\r
1137 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1139 .global DrawAllSprites @ int prio, int sh
\r
1142 ldr r3, =rendstatus
\r
1143 orr r1, r1, r0, lsl #1
\r
1145 tst r12,#(PDRAW_ACC_SPRITES|PDRAW_SPRITES_MOVED)
\r
1147 stmfd sp!, {r1,lr}
\r
1148 and r0, r12,#PDRAW_DIRTY_SPRITES
\r
1149 bic r12,r12,#(PDRAW_ACC_SPRITES|PDRAW_SPRITES_MOVED)
\r
1152 ldmfd sp!, {r1,lr}
\r
1155 ldr r3, =DrawScanline
\r
1156 ldr r2, =HighLnSpr
\r
1158 mov r3, #(MAX_LINE_SPRITES+2)
\r
1159 mla r2, r12, r3, r2
\r
1161 ands r3, r3, #0x7f
\r
1164 @ time to do some real work
\r
1165 stmfd sp!, {r4-r11,lr}
\r
1166 ldr r4, =rendstatus
\r
1169 strb r12,[r2,#1] @ set end marker
\r
1171 add r10,r10,r3 @ r10=HighLnSpr end
\r
1173 mov r8, r1, lsl #4
\r
1174 tst r4, #PDRAW_ACC_SPRITES
\r
1176 str r8, [sp, #-4] @ no calls after this point
\r
1178 .if OVERRIDE_HIGHCOL
\r
1186 ldr lr, =(Pico+0x10000) @ lr=Pico.vram
\r
1188 @ + 0 : hhhhvvvv ----hhvv yyyyyyyy yyyyyyyy // v, h: horiz. size
\r
1189 @ + 4 : xxxxxxxx xxxxxxxx pccvhnnn nnnnnnnn // x: x coord + 8
\r
1191 DrawSprite: @ was: unsigned int *sprite, int sh, int acc_sprites
\r
1192 @ draw next sprite
\r
1193 ldrb r0, [r10,#-1]!
\r
1194 ldr r1, =HighPreSpr
\r
1196 mov r2, r0, lsr #7
\r
1198 ldmeqfd sp!, {r4-r11,pc} @ end of list
\r
1199 cmp r2, r8, lsr #5
\r
1200 bne DrawSprite @ wrong priority
\r
1202 add r0, r1, r0, lsl #3
\r
1204 @ stmfd sp!, {r4-r9,r11,lr}
\r
1205 @ orr r8, r2, r1, lsl #4
\r
1207 ldr r3, [r0] @ sprite[0]
\r
1208 ldr r7, =DrawScanline
\r
1209 mov r6, r3, lsr #28
\r
1210 sub r6, r6, #1 @ r6=width-1 (inc later)
\r
1211 mov r5, r3, lsr #24
\r
1212 and r5, r5, #7 @ r5=height
\r
1214 mov r4, r3, lsl #16 @ r4=sy<<16 (tmp)
\r
1218 sub r7, r7, r4, asr #16 @ r7=row=DrawScanline-sy
\r
1220 mov r2, r9, asr #16 @ r2=sx
\r
1221 mov r9, r9, lsl #16
\r
1222 mov r9, r9, lsr #16
\r
1223 orr r9, r9, r8, lsl #27 @ r9=code|sh[31]|as[27]
\r
1226 movne r4, r5, lsl #3
\r
1228 subne r7, r4, r7 @ if (code&0x1000) row=(height<<3)-1-row; // Flip Y
\r
1230 add r8, r9, r7, lsr #3 @ tile+=row>>3; // Tile number increases going down
\r
1232 mlane r8, r5, r6, r8 @ if (code&0x0800) { tile+=delta*(width-1);
\r
1233 rsbne r5, r5, #0 @ delta=-delta; } // r5=delta now
\r
1235 mov r8, r8, lsl #21
\r
1236 mov r8, r8, lsr #17
\r
1238 add r8, r8, r7, lsl #1 @ tile+=(row&7)<<1; // Tile address
\r
1241 @ cache some stuff to avoid mem access
\r
1242 mov r5, r5, lsl #4 @ delta<<=4; // Delta of address
\r
1243 and r4, r9, #0x6000
\r
1244 orr r9, r9, r4, lsl #16
\r
1245 orr r9, r9, #0x10000000 @ r9=scc1 a??? ... <code> (s=shadow/hilight, cc=pal, a=acc_spr)
\r
1247 orrs r3, r9, r9, lsl #4
\r
1248 mov r3, r4, lsr #9 @ r3=pal=((code>>9)&0x30);
\r
1249 orrmi r3, r3, #0x40 @ for shadow|as
\r
1251 add r6, r6, #1 @ inc now
\r
1252 adds r0, r2, #0 @ mov sx to r0 and set ZV flags
\r
1253 b .dspr_loop_enter
\r
1256 subs r6, r6, #1 @ width--
\r
1258 adds r0, r0, #8 @ sx+=8
\r
1259 add r8, r8, r5 @ tile+=delta
\r
1262 ble .dspr_loop @ sx <= 0
\r
1266 mov r8, r8, lsl #17
\r
1267 mov r8, r8, lsr #17 @ tile&=0x7fff; // Clip tile address
\r
1269 ldr r2, [lr, r8, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
1270 add r1, r11, r0 @ r1=pdest
\r
1274 cmp r12, r9, lsr #28
\r
1277 cmp r2, r2, ror #4
\r
1278 beq .dspr_SingleColor @ tileline singlecolor
\r
1281 bne .dspr_TileFlip
\r
1283 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
1294 cmp r2, #0xe0000000
\r
1295 bcs .dspr_loop @ operator tileline, ignore
\r
1297 .dspr_SingleColor:
\r
1300 orr r4, r4, r4, lsl #8
\r
1301 tst r0, #1 @ not aligned?
\r
1302 strneb r4, [r1], #1
\r
1303 streqh r4, [r1], #2
\r
1307 strneb r4, [r1], #1
\r
1311 cmp r2, r2, ror #4
\r
1312 beq .dspr_singlec_sh
\r
1315 bne .dspr_TileFlip_sh
\r
1317 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
1318 .dspr_TileNorm_sh:
\r
1322 .dspr_TileFlip_sh:
\r
1327 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1329 .global DrawWindow @ int tstart, int tend, int prio, int sh // int *hcache
\r
1332 stmfd sp!, {r4-r11,lr}
\r
1334 ldr r11, =(Pico+0x22228) @ Pico.video
\r
1335 ldr r10, =DrawScanline
\r
1336 ldrb r12, [r11, #3] @ pvid->reg[3]
\r
1339 ldr r4, [r11, #12]
\r
1340 mov r5, r10, lsr #3
\r
1342 mov r10, r10, lsl #1 @ r10=ty
\r
1344 mov r12, r12, lsl #10
\r
1346 tst r4, #1 @ 40 cell mode?
\r
1347 andne r12, r12, #0xf000 @ 0x3c<<10
\r
1348 andeq r12, r12, #0xf800
\r
1349 addne r12, r12, r5, lsl #7
\r
1350 addeq r12, r12, r5, lsl #6 @ nametab
\r
1351 add r12, r12, r0, lsl #2 @ +starttile
\r
1353 ldr r6, =rendstatus
\r
1354 ldr lr, =(Pico+0x10000) @ lr=Pico.vram
\r
1357 @ fetch the first code now
\r
1358 ldrh r7, [lr, r12]
\r
1360 ands r6, r6, #PDRAW_WND_DIFF_PRIO
\r
1363 eoreq r8, r2, r7, lsr #15 @ do prio bits differ?
\r
1365 ldmeqfd sp!, {r4-r11,pc} @ yes, assume that whole window uses same priority
\r
1367 orr r6, r6, r3, lsl #8 @ shadow mode
\r
1371 @ cache some stuff to avoid mem access
\r
1372 .if OVERRIDE_HIGHCOL
\r
1374 mov r8, r8, lsl #1 @ cells
\r
1376 mvn r9, #0 @ r9=prevcode=-1
\r
1379 ldr r11,=(HighCol+8)
\r
1380 mov r8, r8, lsl #1 @ cells
\r
1381 mvn r9, #0 @ r9=prevcode=-1
\r
1383 add r1, r11, r0, lsl #4 @ r1=pdest
\r
1387 @ r4,r5 are scratch in this loop
\r
1391 add r12, r12, #2 @ halfwords
\r
1392 ldrh r7, [lr, r12] @ r7=code (int, but from unsigned, no sign extend)
\r
1394 beq .dwloop_end @ done
\r
1396 eor r5, r6, r7, lsr #15
\r
1398 orrne r6, r6, #2 @ wrong pri
\r
1402 beq .dw_samecode @ we know stuff about this tile already
\r
1405 mov r9, r7 @ remember code
\r
1407 movs r2, r9, lsl #20 @ if (code&0x1000)
\r
1408 mov r2, r2, lsl #1
\r
1409 add r2, r10, r2, lsr #17 @ r2=addr=(code&0x7ff)<<4; addr+=ty
\r
1410 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
1412 and r3, r9, #0x6000
\r
1413 mov r3, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);
\r
1415 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
1422 beq .dwloop @ tileline blank
\r
1424 cmp r2, r2, ror #4
\r
1425 beq .dw_SingleColor @ tileline singlecolor
\r
1430 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r0: helper pattern
\r
1440 and r4, r0, r2 @ #0x0000000f
\r
1442 orr r4, r4, r4, lsl #8
\r
1443 orr r4, r4, r4, lsl #16
\r
1445 stmia r1!, {r4,r5}
\r
1446 b .dwloop_nor1 @ we incremeted r1 ourselves
\r
1449 tst r6, #1 @ hi pri?
\r
1450 orreq r3, r3, #0x40
\r
1451 beq .dw_shadow_done
\r
1454 orr r5, r5, r5, lsl #8
\r
1455 orr r5, r5, r5, lsl #16
\r
1464 ldr r0, =rendstatus
\r
1470 ldmfd sp!, {r4-r11,r12}
\r
1474 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1477 @ hilights 2 pixels in RGB444/BGR444 format
\r
1478 .macro TileDoShHi2Pixels444 reg
\r
1479 mov \reg, \reg, ror #12
\r
1480 adds \reg, \reg, #0x40000000
\r
1481 orrcs \reg, \reg, #0xf0000000
\r
1482 mov \reg, \reg, ror #28
\r
1483 adds \reg, \reg, #0x40000000
\r
1484 orrcs \reg, \reg, #0xf0000000
\r
1485 mov \reg, \reg, ror #28
\r
1486 adds \reg, \reg, #0x40000000
\r
1487 orrcs \reg, \reg, #0xf0000000
\r
1488 mov \reg, \reg, ror #24
\r
1489 adds \reg, \reg, #0x40000000
\r
1490 orrcs \reg, \reg, #0xf0000000
\r
1491 mov \reg, \reg, ror #28
\r
1492 adds \reg, \reg, #0x40000000
\r
1493 orrcs \reg, \reg, #0xf0000000
\r
1494 mov \reg, \reg, ror #28
\r
1495 adds \reg, \reg, #0x40000000
\r
1496 orrcs \reg, \reg, #0xf0000000
\r
1497 mov \reg, \reg, ror #12
\r
1501 .global FinalizeLineBGR444 @ int sh
\r
1503 FinalizeLineBGR444:
\r
1504 stmfd sp!, {r4-r6,lr}
\r
1506 ldr lr, =(Pico+0x22228) @ Pico.video
\r
1507 ldr r0, =DrawLineDest
\r
1508 ldrb r12, [lr, #12]
\r
1510 sub r3, lr, #0x128 @ r3=Pico.cram
\r
1513 movne r2, #320/4 @ len
\r
1514 bne .fl_no32colBGR444
\r
1519 addeq r0, r0, #32*2
\r
1521 .fl_no32colBGR444:
\r
1523 beq .fl_noshBGR444
\r
1527 ldrb r12, [lr, #-0x1a] @ 0x2220e ~ dirtyPal
\r
1530 beq .fl_noshBGR444
\r
1532 strb r12, [lr, #-0x1a]
\r
1537 ldmia r3!, {r1,r5,r6,r12}
\r
1539 stmia r4!, {r1,r5,r6,r12}
\r
1540 bne .fl_loopcpBGR444
\r
1542 @ shadowed pixels:
\r
1544 orr r12,r12,#0x0700
\r
1545 orr r12,r12,r12,lsl #16
\r
1546 sub r3, r3, #0x40*2
\r
1547 add r5, r4, #0x80*2
\r
1549 .fl_loopcpBGR444_sh:
\r
1550 ldmia r3!, {r1,r6}
\r
1552 and r1, r12, r1, lsr #1
\r
1553 and r6, r12, r6, lsr #1
\r
1554 stmia r4!, {r1,r6}
\r
1555 stmia r5!, {r1,r6}
\r
1556 bne .fl_loopcpBGR444_sh
\r
1558 @ hilighted pixels:
\r
1559 sub r3, r3, #0x40*2
\r
1561 .fl_loopcpBGR444_hi:
\r
1563 TileDoShHi2Pixels444 r1
\r
1566 bne .fl_loopcpBGR444_hi
\r
1568 sub r3, r4, #0x40*3*2
\r
1573 ldr r12,=rendstatus
\r
1574 eors r6, r6, #1 @ sh is 0
\r
1577 tstne r12,#PDRAW_ACC_SPRITES
\r
1579 .if OVERRIDE_HIGHCOL
\r
1583 mov lr, lr, lsl #1
\r
1586 ldr r1, =(HighCol+8)
\r
1588 mov lr, lr, lsl #1
\r
1595 and r4, lr, r12, lsl #1
\r
1597 and r5, lr, r12, lsr #7
\r
1599 and r6, lr, r12, lsr #15
\r
1601 and r12,lr, r12, lsr #23
\r
1602 ldrh r12,[r3, r12] @ 1c.i.
\r
1603 orr r4, r4, r5, lsl #16
\r
1604 orr r5, r6, r12,lsl #16
\r
1606 stmia r0!, {r4,r5}
\r
1607 bne .fl_loopBGR444
\r
1610 ldmfd sp!, {r4-r6,lr}
\r
1614 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1617 @ hilights 2 pixels in RGB555/BGR555 format
\r
1618 .macro TileDoShHi2Pixels555 reg
\r
1619 adds \reg, \reg, #0x40000000
\r
1620 orrcs \reg, \reg, #0xf0000000
\r
1621 mov \reg, \reg, ror #27
\r
1622 adds \reg, \reg, #0x40000000
\r
1623 orrcs \reg, \reg, #0xf0000000
\r
1624 mov \reg, \reg, ror #26
\r
1625 adds \reg, \reg, #0x40000000
\r
1626 orrcs \reg, \reg, #0xf0000000
\r
1627 mov \reg, \reg, ror #27
\r
1628 adds \reg, \reg, #0x40000000
\r
1629 orrcs \reg, \reg, #0xf0000000
\r
1630 mov \reg, \reg, ror #27
\r
1631 adds \reg, \reg, #0x40000000
\r
1632 orrcs \reg, \reg, #0xf0000000
\r
1633 mov \reg, \reg, ror #26
\r
1634 adds \reg, \reg, #0x40000000
\r
1635 orrcs \reg, \reg, #0xf0000000
\r
1636 mov \reg, \reg, ror #27
\r
1640 @ Convert 0000bbb0 ggg0rrr0
\r
1641 @ to rrrrrggg gggbbbbb
\r
1643 @ r2,r3,r9 - scratch, lr = 0x001c001c, r8 = 0x00030003
\r
1644 .macro convRGB565 reg
\r
1645 and r2, lr, \reg,lsl #1
\r
1646 and r9, r8, \reg,lsr #2
\r
1647 orr r2, r2, r9 @ r2=red
\r
1648 and r3, lr, \reg,lsr #7
\r
1649 and r9, r8, \reg,lsr #10
\r
1650 orr r3, r3, r9 @ r3=blue
\r
1651 and \reg, \reg, lr, lsl #3
\r
1652 orr \reg, \reg, \reg,lsl #3 @ green
\r
1653 orr \reg, \reg, r2, lsl #11 @ add red back
\r
1654 orr \reg, \reg, r3 @ add blue back
\r
1657 .global vidConvCpyRGB565
\r
1659 vidConvCpyRGB565: @ void *to, void *from, int pixels
\r
1660 stmfd sp!, {r4-r9,lr}
\r
1662 mov r12, r2, lsr #3 @ repeats
\r
1663 mov lr, #0x001c0000
\r
1664 orr lr, lr, #0x01c @ lr == pattern 0x001c001c
\r
1665 mov r8, #0x00030000
\r
1666 orr r8, r8, #0x003
\r
1669 ldmia r1!, {r4-r7}
\r
1682 ldmfd sp!, {r4-r9,lr}
\r
1687 .global FinalizeLineRGB555 @ int sh
\r
1689 FinalizeLineRGB555:
\r
1690 stmfd sp!, {r4-r8,lr}
\r
1691 ldr r8, =(Pico+0x22228) @ Pico.video
\r
1694 ldrb r7, [r8, #-0x1a] @ 0x2220e ~ dirtyPal
\r
1698 beq .fl_noconvRGB555
\r
1699 strb r1, [r8, #-0x1a]
\r
1700 sub r1, r8, #0x128 @ r1=Pico.cram
\r
1703 bl vidConvCpyRGB565
\r
1708 beq .fl_noshRGB555
\r
1710 beq .fl_noshRGB555
\r
1712 @ shadowed pixels:
\r
1714 orr r12,r12,#0x7300
\r
1715 orr r12,r12,r12,lsl #16
\r
1716 add r4, r3, #0x40*2
\r
1717 add r5, r3, #0xc0*2
\r
1719 .fl_loopcpRGB555_sh:
\r
1720 ldmia r3!, {r1,r6}
\r
1722 and r1, r12, r1, lsr #1
\r
1723 and r6, r12, r6, lsr #1
\r
1724 stmia r4!, {r1,r6}
\r
1725 stmia r5!, {r1,r6}
\r
1726 bne .fl_loopcpRGB555_sh
\r
1728 @ hilighted pixels:
\r
1729 sub r3, r3, #0x40*2
\r
1731 .fl_loopcpRGB555_hi:
\r
1733 TileDoShHi2Pixels555 r1
\r
1736 bne .fl_loopcpRGB555_hi
\r
1738 sub r3, r3, #0x40*2
\r
1742 ldr r12,=rendstatus
\r
1743 eors r6, r6, #1 @ sh is 0
\r
1746 tstne r12,#PDRAW_ACC_SPRITES
\r
1749 .if OVERRIDE_HIGHCOL
\r
1751 ldr r0, =DrawLineDest
\r
1756 ldr r0, =DrawLineDest
\r
1757 ldr r1, =(HighCol+8)
\r
1761 ldrb r12, [r8, #12]
\r
1762 mov lr, lr, lsl #1
\r
1765 movne r2, #320/8 @ len
\r
1766 bne .fl_no32colRGB555
\r
1771 bne .fl_32scale_RGB555
\r
1773 addeq r0, r0, #32*2
\r
1775 .fl_no32colRGB555:
\r
1777 .if UNALIGNED_DRAWLINEDEST
\r
1778 @ this is basically for Gizmondo, which has unaligned odd lines in the framebuffer
\r
1787 and r4, lr, r12, lsl #1
\r
1789 and r5, lr, r12, lsr #7
\r
1791 and r6, lr, r12, lsr #15
\r
1793 orr r4, r4, r5, lsl #16
\r
1795 and r5, lr, r12, lsr #23
\r
1797 and r8, lr, r7, lsl #1
\r
1799 orr r5, r6, r5, lsl #16
\r
1801 and r6, lr, r7, lsr #7
\r
1803 and r12,lr, r7, lsr #15
\r
1804 ldrh r12,[r3, r12]
\r
1805 and r7, lr, r7, lsr #23
\r
1807 orr r8, r8, r6, lsl #16
\r
1810 orr r12,r12, r7, lsl #16
\r
1812 stmia r0!, {r4,r5,r8,r12}
\r
1813 bne .fl_loopRGB555
\r
1815 ldmfd sp!, {r4-r8,lr}
\r
1819 .fl_32scale_RGB555:
\r
1820 stmfd sp!, {r9,r10}
\r
1821 mov r9, #0x3900 @ f800 07e0 001f | e000 0780 001c | 3800 01e0 0007
\r
1822 orr r9, r9, #0x00e7
\r
1824 .if UNALIGNED_DRAWLINEDEST
\r
1826 bne .fl_32scale_RGB555u
\r
1829 .fl_loop32scale_RGB555:
\r
1833 and r4, lr, r12,lsl #1
\r
1835 and r5, lr, r12,lsr #7
\r
1837 and r4, r4, r9, lsl #2
\r
1838 orr r4, r4, r4, lsl #14 @ r4[31:16] = 1/4 pix_s 0
\r
1839 and r5, r5, r9, lsl #2
\r
1840 sub r6, r5, r5, lsr #2 @ r6 = 3/4 pix_s 1
\r
1841 add r4, r4, r6, lsl #16 @ pix_d 0, 1
\r
1842 and r6, lr, r12,lsr #15
\r
1844 and r12,lr, r12,lsr #23
\r
1845 ldrh r12,[r3, r12]
\r
1846 and r6, r6, r9, lsl #2
\r
1848 mov r5, r5, lsr #1
\r
1849 sub r6, r6, r6, lsr #2 @ r6 = 3/4 pix_s 2
\r
1850 orr r5, r5, r6, lsl #16
\r
1852 and r6, lr, r7, lsl #1
\r
1854 and r12,r12,r9, lsl #2
\r
1855 add r5, r5, r12,lsl #14 @ pix_d 2, 3
\r
1856 and r6, r6, r9, lsl #2
\r
1857 orr r6, r12,r6, lsl #16 @ pix_d 4, 5
\r
1859 and r12,lr, r7, lsr #7
\r
1860 ldrh r12,[r3, r12]
\r
1861 and r10,lr, r7, lsr #15
\r
1862 ldrh r10,[r3, r10]
\r
1863 and r12,r12,r9, lsl #2
\r
1864 sub r8, r12,r12,lsr #2 @ r8 = 3/4 pix_s 1
\r
1865 add r8, r8, r6, lsr #18
\r
1866 and r7, lr, r7, lsr #23
\r
1868 and r10,r10,r9, lsl #2
\r
1869 orr r8, r8, r10,lsl #15
\r
1870 add r8, r8, r12,lsl #15 @ pix_d 6, 7
\r
1871 sub r10,r10,r10,lsr #2 @ r10= 3/4 pix_s 2
\r
1872 and r7, r7, r9, lsl #2
\r
1873 add r10,r10,r7, lsr #2 @ += 1/4 pix_s 3
\r
1874 orr r10,r10,r7, lsl #16 @ pix_d 8, 9
\r
1878 stmia r0!, {r4,r5,r6,r8,r10}
\r
1879 bne .fl_loop32scale_RGB555
\r
1881 ldmfd sp!, {r9,r10}
\r
1882 ldmfd sp!, {r4-r8,lr}
\r
1885 .if UNALIGNED_DRAWLINEDEST
\r
1886 @ unaligned versions of loops
\r
1887 @ warning: starts drawing 2bytes before dst
\r
1890 sub r0, r0, #2 @ initial adjustment
\r
1897 and r6, lr, r12,lsl #1
\r
1899 and r5, lr, r12,lsr #7
\r
1901 orr r4, r8, r6, lsl #16
\r
1903 and r6, lr, r12,lsr #15
\r
1905 and r8, lr, r12,lsr #23
\r
1907 orr r5, r5, r6, lsl #16
\r
1909 and r6, lr, r7, lsl #1
\r
1911 and r12,lr, r7, lsr #7
\r
1912 ldrh r12,[r3, r12]
\r
1913 orr r6, r8, r6, lsl #16
\r
1915 and r8, lr, r7, lsr #15
\r
1917 and r7, lr, r7, lsr #23
\r
1920 orr r12,r12,r8, lsl #16
\r
1923 stmia r0!, {r4,r5,r6,r12}
\r
1924 bne .fl_loopRGB555u
\r
1928 ldmfd sp!, {r4-r8,lr}
\r
1932 .fl_32scale_RGB555u:
\r
1933 sub r0, r0, #2 @ initial adjustment
\r
1936 @ r9 f800 07e0 001f | e000 0780 001c | 3800 01e0 0007
\r
1937 .fl_loop32scale_RGB555u:
\r
1941 and r6, lr, r12,lsl #1
\r
1943 and r5, lr, r12,lsr #7
\r
1945 and r6, r6, r9, lsl #2
\r
1946 orr r4, r4, r6, lsl #16 @ r4 = pix_d -1, 0
\r
1948 and r5, r5, r9, lsl #2
\r
1949 sub r8, r5, r5, lsr #2 @ r8 = 3/4 pix_s 1
\r
1950 add r6, r8, r6, lsr #2 @ r6 = (1/4 pix_s 0) + (3/4 pix_s 1)
\r
1951 orr r5, r6, r5, lsl #15
\r
1953 and r6, lr, r12,lsr #15
\r
1955 and r12,lr, r12,lsr #23
\r
1956 ldrh r12,[r3, r12]
\r
1957 and r6, r6, r9, lsl #2
\r
1958 add r5, r5, r6, lsl #15 @ r5 = pix_d 1, 2
\r
1960 and r8, lr, r7, lsl #1
\r
1962 and r10,lr, r7, lsr #7
\r
1963 ldrh r10,[r3, r10]
\r
1964 and r12,r12,r9, lsl #2
\r
1965 sub r6, r6, r6, lsr #2 @ r6 = 3/4 pix_s 2
\r
1966 add r6, r6, r12,lsr #2
\r
1967 orr r6, r6, r12,lsl #16 @ r6 = pix_d 3, 4
\r
1969 and r8, r8, r9, lsl #2
\r
1970 and r10,r10,r9, lsl #2
\r
1971 sub r12,r10,r10,lsr #2 @ r12 = 3/4 pix_s 5
\r
1972 orr r8, r8, r8, lsl #14
\r
1973 add r8, r8, r12,lsl #16 @ r8 = pix_d 5, 6
\r
1974 and r12,lr, r7, lsr #15
\r
1975 ldrh r12,[r3, r12]
\r
1976 and r7, lr, r7, lsr #23
\r
1978 and r12,r12,r9, lsl #2
\r
1980 mov r10,r10, lsr #1
\r
1981 sub r12,r12,r12,lsr #2 @ r12 = 3/4 pix_s 6
\r
1982 orr r10,r10,r12,lsl #16
\r
1983 and r7, r7, r9, lsl #2
\r
1984 add r10,r10,r7, lsl #14 @ r10 = pix_d 7, 8
\r
1988 stmia r0!, {r4,r5,r6,r8,r10}
\r
1990 bne .fl_loop32scale_RGB555u
\r
1994 ldmfd sp!, {r9,r10}
\r
1995 ldmfd sp!, {r4-r8,lr}
\r
1998 .endif @ UNALIGNED_DRAWLINEDEST
\r
2001 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
2004 .global blockcpy @ void *dst, void *src, size_t n
\r
2007 stmfd sp!, {r4,r5}
\r
2008 mov r2, r2, lsr #4
\r
2010 ldmia r1!, {r3-r5,r12}
\r
2012 stmia r0!, {r3-r5,r12}
\r
2014 ldmfd sp!, {r4,r5}
\r
2018 .global blockcpy_or @ void *dst, void *src, size_t n, int pat
\r
2021 stmfd sp!, {r4-r6}
\r
2022 orr r3, r3, r3, lsl #8
\r
2023 orr r3, r3, r3, lsl #16
\r
2024 mov r2, r2, lsr #4
\r
2026 ldmia r1!, {r4-r6,r12}
\r
2032 stmia r0!, {r4-r6,r12}
\r
2033 bne blockcpy_loop_or
\r
2034 ldmfd sp!, {r4-r6}
\r