2 * assembly optimized versions of most funtions from draw.c
\r
3 * (C) notaz, 2006-2010
\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
14 .extern DrawScanline
\r
18 .extern DrawLineDest
\r
19 .extern DrawStripInterlace
\r
20 .extern HighCacheS_ptr
\r
22 .equiv OVERRIDE_HIGHCOL, 1
\r
24 .equ PDRAW_SPRITES_MOVED, (1<<0)
\r
25 .equ PDRAW_WND_DIFF_PRIO, (1<<1)
\r
26 .equ PDRAW_ACC_SPRITES, (1<<2)
\r
27 .equ PDRAW_DIRTY_SPRITES, (1<<4)
\r
28 .equ PDRAW_PLANE_HI_PRIO, (1<<6)
\r
29 .equ PDRAW_SHHI_DONE, (1<<7)
\r
32 .macro TilePixel pat lsrr offs
\r
36 ands r4, \pat, r2, lsr #\lsrr
\r
39 strneb r4, [r1,#\offs]
\r
42 @ TileNorm (r1=pdest, r2=pixels8, r3=pal) r4: scratch, pat: register with helper pattern 0xf
\r
44 TilePixel \pat, 12, 0 @ #0x0000f000
\r
45 TilePixel \pat, 8, 1 @ #0x00000f00
\r
46 TilePixel \pat, 4, 2 @ #0x000000f0
\r
47 TilePixel \pat, 0, 3 @ #0x0000000f
\r
48 TilePixel \pat, 28, 4 @ #0xf0000000
\r
49 TilePixel \pat, 24, 5 @ #0x0f000000
\r
50 TilePixel \pat, 20, 6 @ #0x00f00000
\r
51 TilePixel \pat, 16, 7 @ #0x000f0000
\r
54 @ TileFlip (r1=pdest, r2=pixels8, r3=pal) r4: scratch, pat: register with helper pattern 0xf
\r
56 TilePixel \pat, 16, 0 @ #0x000f0000
\r
57 TilePixel \pat, 20, 1 @ #0x00f00000
\r
58 TilePixel \pat, 24, 2 @ #0x0f000000
\r
59 TilePixel \pat, 28, 3 @ #0xf0000000
\r
60 TilePixel \pat, 0, 4 @ #0x0000000f
\r
61 TilePixel \pat, 4, 5 @ #0x000000f0
\r
62 TilePixel \pat, 8, 6 @ #0x00000f00
\r
63 TilePixel \pat, 12, 7 @ #0x0000f000
\r
66 @ shadow/hilight mode
\r
68 @ this one is for hi priority layer
\r
69 .macro TilePixelShHP lsrr offs
\r
73 ands r4, r12, r2, lsr #\lsrr
\r
75 ldreqb r4, [r1,#\offs]
\r
78 strb r4, [r1,#\offs]
\r
81 @ TileNormShHP (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: register with helper pattern 0xf, touches r3 high bits
\r
83 TilePixelShHP 12, 0 @ #0x0000f000
\r
84 TilePixelShHP 8, 1 @ #0x00000f00
\r
85 TilePixelShHP 4, 2 @ #0x000000f0
\r
86 TilePixelShHP 0, 3 @ #0x0000000f
\r
87 TilePixelShHP 28, 4 @ #0xf0000000
\r
88 TilePixelShHP 24, 5 @ #0x0f000000
\r
89 TilePixelShHP 20, 6 @ #0x00f00000
\r
90 TilePixelShHP 16, 7 @ #0x000f0000
\r
93 @ TileFlipShHP (r1=pdest, r2=pixels8, r3=pal) r4: scratch, pat: register with helper pattern 0xf
\r
95 TilePixelShHP 16, 0 @ #0x000f0000
\r
96 TilePixelShHP 20, 1 @ #0x00f00000
\r
97 TilePixelShHP 24, 2 @ #0x0f000000
\r
98 TilePixelShHP 28, 3 @ #0xf0000000
\r
99 TilePixelShHP 0, 4 @ #0x0000000f
\r
100 TilePixelShHP 4, 5 @ #0x000000f0
\r
101 TilePixelShHP 8, 6 @ #0x00000f00
\r
102 TilePixelShHP 12, 7 @ #0x0000f000
\r
106 @ TileSingleSh (r1=pdest, r2=pixels8, r3=pal) r4,r7: scratch, r0=sx; r12: helper pattern 0xf
\r
107 .macro TileSingleSh
\r
108 tst r0, #1 @ not aligned?
\r
114 strneb r4, [r1], #1
\r
115 streqh r4, [r1], #2
\r
127 strneb r4, [r1], #1
\r
130 @ TileSingleHi (r1=pdest, r2=pixels8, r3=pal) r4,r7: scratch, r0=sx, r12: register with helper pattern 0xf
\r
131 .macro TileSingleHi
\r
132 tst r1, #1 @ not aligned?
\r
135 ldrneb r4, [r1], #1
\r
136 ldreqh r4, [r1], #2 @ 1ci
\r
138 bic r4, r4, r7, lsr #1
\r
140 strneb r4, [r1, #-3]
\r
141 streqh r4, [r1, #-4]
\r
143 bic r12, r12, r7, lsr #1
\r
145 strh r12, [r1, #-4]
\r
147 bic r4, r4, r7, lsr #1
\r
151 bic r12, r12, r7, lsr #1
\r
153 strh r12, [r1, #-2]
\r
154 bicne r4, r4, r7, lsr #1
\r
156 strneb r4, [r1], #1
\r
160 .macro TileDoShGenPixel shift ofs
\r
162 ands r4, r12, r2, lsr #\shift
\r
168 ldrgeb r7, [r1,#\ofs]
\r
169 orrlt r7, r3, r4 @ normal
\r
171 bicge r7, r7, #0xc0
\r
172 orrge r7, r7, r4, lsl #6
\r
173 strb r7, [r1,#\ofs]
\r
177 @ TileFlipSh (r1=pdest, r2=pixels8, r3=pal) r4,r7: scratch, r0=sx, r12: register with helper pattern 0xf
\r
179 TileDoShGenPixel 16, 0 @ #0x000f0000
\r
180 TileDoShGenPixel 20, 1 @ #0x00f00000
\r
181 TileDoShGenPixel 24, 2 @ #0x0f000000
\r
182 TileDoShGenPixel 28, 3 @ #0xf0000000
\r
183 TileDoShGenPixel 0, 4 @ #0x0000000f
\r
184 TileDoShGenPixel 4, 5 @ #0x000000f0
\r
185 TileDoShGenPixel 8, 6 @ #0x00000f00
\r
186 TileDoShGenPixel 12, 7 @ #0x0000f000
\r
189 @ TileNormSh (r1=pdest, r2=pixels8, r3=pal) r4,r7: scratch, r0=sx, r12: register with helper pattern 0xf
\r
191 TileDoShGenPixel 12, 0 @ #0x0000f000
\r
192 TileDoShGenPixel 8, 1 @ #0x00000f00
\r
193 TileDoShGenPixel 4, 2 @ #0x000000f0
\r
194 TileDoShGenPixel 0, 3 @ #0x0000000f
\r
195 TileDoShGenPixel 28, 4 @ #0xf0000000
\r
196 TileDoShGenPixel 24, 5 @ #0x0f000000
\r
197 TileDoShGenPixel 20, 6 @ #0x00f00000
\r
198 TileDoShGenPixel 16, 7 @ #0x000f0000
\r
201 .macro TileDoShGenPixel_markop shift ofs
\r
203 ands r4, r12, r2, lsr #\shift
\r
209 ldrgeb r4, [r1,#\ofs]
\r
211 orrge r4, r4, #0x80
\r
212 strb r4, [r1,#\ofs]
\r
216 .macro TileFlipSh_markop
\r
217 TileDoShGenPixel_markop 16, 0 @ #0x000f0000
\r
218 TileDoShGenPixel_markop 20, 1 @ #0x00f00000
\r
219 TileDoShGenPixel_markop 24, 2 @ #0x0f000000
\r
220 TileDoShGenPixel_markop 28, 3 @ #0xf0000000
\r
221 TileDoShGenPixel_markop 0, 4 @ #0x0000000f
\r
222 TileDoShGenPixel_markop 4, 5 @ #0x000000f0
\r
223 TileDoShGenPixel_markop 8, 6 @ #0x00000f00
\r
224 TileDoShGenPixel_markop 12, 7 @ #0x0000f000
\r
227 .macro TileNormSh_markop
\r
228 TileDoShGenPixel_markop 12, 0 @ #0x0000f000
\r
229 TileDoShGenPixel_markop 8, 1 @ #0x00000f00
\r
230 TileDoShGenPixel_markop 4, 2 @ #0x000000f0
\r
231 TileDoShGenPixel_markop 0, 3 @ #0x0000000f
\r
232 TileDoShGenPixel_markop 28, 4 @ #0xf0000000
\r
233 TileDoShGenPixel_markop 24, 5 @ #0x0f000000
\r
234 TileDoShGenPixel_markop 20, 6 @ #0x00f00000
\r
235 TileDoShGenPixel_markop 16, 7 @ #0x000f0000
\r
238 .macro TileDoShGenPixel_onlyop_lp shift ofs
\r
240 ands r7, r12, r2, lsr #\shift
\r
244 ldrneb r4, [r1,#\ofs]
\r
249 bicne r4, r4, #0xc0
\r
250 orrne r4, r4, r7, lsl #6
\r
251 strneb r4, [r1,#\ofs]
\r
255 .macro TileFlipSh_onlyop_lp
\r
256 TileDoShGenPixel_onlyop_lp 16, 0 @ #0x000f0000
\r
257 TileDoShGenPixel_onlyop_lp 20, 1 @ #0x00f00000
\r
258 TileDoShGenPixel_onlyop_lp 24, 2 @ #0x0f000000
\r
259 TileDoShGenPixel_onlyop_lp 28, 3 @ #0xf0000000
\r
260 TileDoShGenPixel_onlyop_lp 0, 4 @ #0x0000000f
\r
261 TileDoShGenPixel_onlyop_lp 4, 5 @ #0x000000f0
\r
262 TileDoShGenPixel_onlyop_lp 8, 6 @ #0x00000f00
\r
263 TileDoShGenPixel_onlyop_lp 12, 7 @ #0x0000f000
\r
266 .macro TileNormSh_onlyop_lp
\r
267 TileDoShGenPixel_onlyop_lp 12, 0 @ #0x0000f000
\r
268 TileDoShGenPixel_onlyop_lp 8, 1 @ #0x00000f00
\r
269 TileDoShGenPixel_onlyop_lp 4, 2 @ #0x000000f0
\r
270 TileDoShGenPixel_onlyop_lp 0, 3 @ #0x0000000f
\r
271 TileDoShGenPixel_onlyop_lp 28, 4 @ #0xf0000000
\r
272 TileDoShGenPixel_onlyop_lp 24, 5 @ #0x0f000000
\r
273 TileDoShGenPixel_onlyop_lp 20, 6 @ #0x00f00000
\r
274 TileDoShGenPixel_onlyop_lp 16, 7 @ #0x000f0000
\r
278 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
282 @ int nametab; // 0x00
\r
283 @ int line; // 0x04
\r
284 @ int hscroll; // 0x08
\r
285 @ int xmask; // 0x0C
\r
286 @ int *hc; // 0x10 (pointer to cache buffer)
\r
287 @ int cells; // 0x14
\r
290 @ void DrawLayer(int plane_sh, int *hcache, int cellskip, int maxcells);
\r
295 stmfd sp!, {r4-r11,lr}
\r
297 ldr r11, =(Pico+0x22228) @ Pico.video
\r
300 ldrb r7, [r11, #16] @ ??vv??hh
\r
302 mov r6, r1 @ hcache
\r
303 orr r9, r3, r0, lsl #30
\r
304 orr r9, r9, r2, lsl #8 @ r9=sh[31]|cellskip[15:8]|maxcells[7:0] (tmp)
\r
307 orr r1, r1, #0x00ff
\r
311 biclt r1, r1, #0xfc00
\r
312 biceq r1, r1, #0xfe00
\r
313 bicgt r1, r1, #0xff00 @ r1=ymask=(height<<8)|0xff; ...; // Y Mask in pixels
\r
317 subge r10, r10, #1 @ r10=shift[width] (5,6,6,7)
\r
320 mov r5, r8, lsl r10
\r
321 sub r5, r5, #1 @ r5=xmask
\r
325 ldreqb r12, [r11, #2]
\r
326 ldrneb r12, [r11, #4]
\r
328 ldr r2, =DrawScanline @ trying to make good use of pipeline here
\r
329 ldr lr, =(Pico+0x10000) @ lr=Pico.vram
\r
331 moveq r12, r12, lsl #10
\r
332 movne r12, r12, lsl #13
\r
333 and r12, r12, #(7<<13) @ r12=(ts->nametab<<1) (halfword compliant)
\r
335 ldrh r8, [r11, #12]
\r
336 ldrb r7, [r11, #11]
\r
339 mov r4, r8, lsr #8 @ pvid->reg[13]
\r
340 mov r4, r4, lsl #10 @ htab=pvid->reg[13]<<9; (halfwords)
\r
342 addne r4, r4, r2, lsl #2 @ htab+=DrawScanline<<1; // Offset by line
\r
344 biceq r4, r4, #0x1f @ htab&=~0xf; // Offset by tile
\r
345 add r4, r4, r0, lsl #1 @ htab+=plane
\r
346 bic r4, r4, #0x00ff0000 @ just in case
\r
347 ldrh r3, [lr, r4] @ r3=hscroll
\r
350 bne .DrawStrip_vsscroll
\r
352 @ Get vertical scroll value:
\r
353 add r7, lr, #0x012000
\r
354 add r7, r7, #0x000180 @ r7=Pico.vsram (Pico+0x22180)
\r
359 bne .DrawStrip_interlace
\r
362 movne r7, r7, lsr #16
\r
364 @ Find the line in the name table
\r
368 add r10, r10, #1 @ shift[width]++
\r
369 add r12, r12, r4, lsl r10 @ nametab+=(ts.line>>3)<<shift[width];
\r
371 @ ldmia r0, {r1,r2,r3,r5,r6,r9} @ r2=line, r3=ts->hscroll, r5=ts->xmask, r6=ts->hc, r9=ts->cells
\r
374 mov r10,r10, lsl #1 @ r10=ty=(ts->line&7)<<1;
\r
375 orr r10,r10, r9, lsl #24
\r
378 mov r8, r8, lsr #3 @ r8=tilex=(-ts->hscroll)>>3
\r
382 add r7, r1, #1 @ r7=dx=((ts->hscroll-1)&7)+1
\r
386 orrne r10,r10, #1<<23 @ r10=(cells<<24|sh<<23|hi_not_empty<<22|had_output<<21|ty)
\r
387 movne r3, #0x40 @ default to shadowed pal on sh mode
\r
390 addne r10,r10, #0x01000000 @ we will loop cells+1 times if there is scroll
\r
392 and r9, r9, #0xff00
\r
393 add r8, r8, r9, lsr #8 @ tilex+=cellskip
\r
394 add r7, r7, r9, lsr #5 @ dx+=cellskip<<3;
\r
395 sub r10,r10,r9, lsl #16 @ cells-=cellskip
\r
397 @ cache some stuff to avoid mem access
\r
398 .if OVERRIDE_HIGHCOL
\r
407 mvn r9, #0 @ r9=prevcode=-1
\r
408 add r1, r11, r7 @ r1=pdest
\r
411 @ r4 & r7 are scratch in this loop
\r
414 .dsloop: @ 40-41 times
\r
415 subs r10,r10, #0x01000000
\r
420 add r7, lr, r7, lsl #1 @ Pico.vram+((tilex&ts->xmask) as halfwords)
\r
421 ldrh r7, [r7, r12] @ r7=code (int, but from unsigned, no sign extend)
\r
427 bne .DrawStrip_hiprio
\r
430 beq .DrawStrip_samecode @ we know stuff about this tile already
\r
432 mov r9, r7 @ remember code
\r
433 orr r10, r10, #1<<21 @ seen non hi-prio tile
\r
435 movs r2, r9, lsl #20 @ if (code&0x1000)
\r
437 add r2, r2, r10, lsl #17
\r
438 mov r2, r2, lsr #17
\r
439 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
441 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
444 and r3, r9, #0x6000
\r
445 add r3, r7, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);
\r
447 .DrawStrip_samecode:
\r
449 beq .dsloop @ tileline blank
\r
452 beq .DrawStrip_SingleColor @ tileline singlecolor
\r
455 bne .DrawStrip_TileFlip
\r
457 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r0: helper pattern
\r
458 .DrawStrip_TileNorm:
\r
462 .DrawStrip_TileFlip:
\r
466 .DrawStrip_SingleColor:
\r
469 orr r4, r4, r4, lsl #8
\r
470 tst r1, #1 @ not aligned?
\r
471 strneb r4, [r1], #1
\r
472 streqh r4, [r1], #2
\r
476 strneb r4, [r1], #1 @ have a remaining unaligned pixel?
\r
479 .DrawStrip_hiprio_maybempt:
\r
481 beq .dsloop @ must've been empty, otherwise we wouldn't get here
\r
482 movs r2, r7, lsl #20 @ if (code&0x1000)
\r
484 add r2, r2, r10, lsl #17
\r
485 mov r2, r2, lsr #17
\r
486 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
487 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
488 mov r9, r7 @ remember code
\r
491 orr r10, r10, #1<<22
\r
494 tst r10, #0x00c00000
\r
495 beq .DrawStrip_hiprio_maybempt
\r
497 orr r7, r7, r0, lsl #16
\r
498 orr r7, r7, r10, lsl #25 @ (ty<<25)
\r
500 eorne r7, r7, #7<<26 @ if(code&0x1000) cval^=7<<26;
\r
501 str r7, [r6], #4 @ cache hi priority tile
\r
506 tst r10, #1<<21 @ seen non hi-prio tile
\r
507 ldreq r1, =rendstatus
\r
510 str r0, [r6] @ terminate the cache list
\r
511 orreq r2, r2, #PDRAW_PLANE_HI_PRIO @ had a layer with all hi-prio tiles
\r
514 ldmfd sp!, {r4-r11,lr}
\r
517 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
519 .DrawStrip_vsscroll:
\r
521 mov r8, r8, lsr #3 @ r8=tilex=(-ts->hscroll)>>3
\r
522 bic r8, r8, #0x3fc00000
\r
523 orr r8, r8, r5, lsl #25 @ r8=(xmask[31:25]|had_output[24]|tilex[21:0])
\r
525 ldr r4, =DrawScanline
\r
526 orr r5, r1, r10, lsl #24
\r
529 orr r5, r5, r4, lsl #16 @ r5=(shift_width[31:24]|scanline[23:16]|ymask[15:0])
\r
531 add r7, r1, #1 @ r7=dx=((ts->hscroll-1)&7)+1
\r
533 mov r10,r9, lsl #16
\r
535 orrne r10,r10, #0x8000
\r
538 orr r10,r10, #0xff000000 @ will be adjusted on entering loop
\r
539 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
540 movne r3, #0x40 @ default to shadowed pal on sh mode
\r
543 subne r10,r10, #0x01000000 @ have hscroll, start with negative cell
\r
545 and r9, r9, #0xff00
\r
546 add r8, r8, r9, lsr #8 @ tilex+=cellskip
\r
547 add r7, r7, r9, lsr #5 @ dx+=cellskip<<3;
\r
548 add r10,r10,r9, lsl #16 @ cell+=cellskip
\r
550 @ cache some stuff to avoid mem access
\r
551 .if OVERRIDE_HIGHCOL
\r
560 mvn r9, #0 @ r9=prevcode=-1
\r
561 add r1, r11, r7 @ r1=pdest
\r
563 @ r4 & r7 are scratch in this loop
\r
566 .dsloop_vs: @ 40-41 times
\r
567 add r10,r10, #0x01000000
\r
568 and r4, r10, #0x003f0000
\r
569 cmp r4, r10, asr #8
\r
570 ble .dsloop_vs_exit
\r
572 @ calc offset and read tileline code to r7, also calc ty
\r
573 add r7, lr, #0x012000
\r
574 add r7, r7, #0x000180 @ r7=Pico.vsram (Pico+0x22180)
\r
575 add r7, r7, r10,asr #23 @ vsram + ((cell&~1)<<1)
\r
577 tst r10,#0x8000 @ plane1?
\r
579 ldrh r7, [r7] @ r7=vscroll
\r
581 bic r10,r10,#0xff @ clear old ty
\r
582 and r4, r5, #0xff0000 @ scanline
\r
583 add r4, r4, r7, lsl #16 @ ... += vscroll
\r
584 and r4, r4, r5, lsl #16 @ ... &= ymask
\r
585 and r7, r4, #0x70000
\r
586 orr r10,r10,r7, lsr #15 @ new ty
\r
588 mov r4, r4, lsr #19
\r
589 mov r7, r5, lsr #24
\r
590 mov r4, r4, lsl r7 @ nametabadd
\r
592 and r7, r8, r8, lsr #25
\r
593 add r7, lr, r7, lsl #1 @ Pico.vram+((tilex&ts->xmask) as halfwords)
\r
594 add r7, r7, r4, lsl #1
\r
595 ldrh r7, [r7, r12] @ r7=code (int, but from unsigned, no sign extend)
\r
601 bne .DrawStrip_vs_hiprio
\r
604 beq .DrawStrip_vs_samecode @ we know stuff about this tile already
\r
606 mov r9, r7 @ remember code
\r
607 orr r8, r8, #(1<<24)@ seen non hi-prio tile
\r
609 movs r2, r9, lsl #20 @ if (code&0x1000)
\r
611 add r2, r2, r10, lsl #17
\r
612 mov r2, r2, lsr #17
\r
613 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
615 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
618 and r3, r9, #0x6000
\r
619 add r3, r7, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);
\r
621 .DrawStrip_vs_samecode:
\r
623 beq .dsloop_vs @ tileline blank
\r
626 beq .DrawStrip_vs_SingleColor @ tileline singlecolor
\r
629 bne .DrawStrip_vs_TileFlip
\r
631 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r0: helper pattern
\r
632 .DrawStrip_vs_TileNorm:
\r
636 .DrawStrip_vs_TileFlip:
\r
640 .DrawStrip_vs_SingleColor:
\r
643 orr r4, r4, r4, lsl #8
\r
644 tst r1, #1 @ not aligned?
\r
645 strneb r4, [r1], #1
\r
646 streqh r4, [r1], #2
\r
650 strneb r4, [r1], #1 @ have a remaining unaligned pixel?
\r
653 .DrawStrip_vs_hiprio:
\r
654 tst r10, #0x00c00000
\r
655 beq .DrawStrip_vs_hiprio_maybempt
\r
657 orr r7, r7, r0, lsl #16
\r
658 orr r7, r7, r10, lsl #25 @ (ty<<25)
\r
660 eorne r7, r7, #7<<26 @ if(code&0x1000) cval^=7<<26;
\r
661 str r7, [r6], #4 @ cache hi priority tile
\r
665 .DrawStrip_vs_hiprio_maybempt:
\r
667 beq .dsloop_vs @ must've been empty, otherwise we wouldn't get here
\r
668 movs r2, r7, lsl #20 @ if (code&0x1000)
\r
670 add r2, r2, r10, lsl #17
\r
671 mov r2, r2, lsr #17
\r
672 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
673 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
674 mov r9, r7 @ remember code
\r
676 orrne r10, r10, #1<<22
\r
677 bne .DrawStrip_vs_hiprio
\r
681 tst r8, #(1<<24) @ seen non hi-prio tile
\r
682 ldreq r1, =rendstatus
\r
685 str r0, [r6] @ terminate the cache list
\r
686 orreq r2, r2, #PDRAW_PLANE_HI_PRIO @ had a layer with all hi-prio tiles
\r
689 ldmfd sp!, {r4-r11,lr}
\r
693 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
695 @ interlace mode 2? Sonic 2?
\r
696 .DrawStrip_interlace:
\r
698 moveq r7, r7, lsl #21
\r
699 movne r7, r7, lsl #5
\r
701 @ Find the line in the name table
\r
702 add r2, r7, r2, lsl #22 @ r2=(vscroll+(DrawScanline<<1))<<21 (11 bits);
\r
703 orr r1, r1, #0x80000000
\r
704 and r2, r2, r1, ror #10 @ &((ymask<<1)|1)<<21;
\r
705 mov r2, r2, lsr #21
\r
707 mov r12, r12, lsr #1 @ halfwords
\r
708 add r0, r12, r4, lsl r10 @ nametab+=(ts.line>>4)<<shift[width];
\r
712 stmia sp, {r0,r2,r3,r5,r6,r9}
\r
715 bl DrawStripInterlace @ struct TileStrip *ts
\r
718 ldmfd sp!, {r4-r11,lr}
\r
723 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
726 .global BackFill @ int reg7, int sh
\r
729 stmfd sp!, {r4-r9,lr}
\r
731 .if OVERRIDE_HIGHCOL
\r
733 mov r0, r0, lsl #26
\r
735 mov r0, r0, lsr #26
\r
738 ldr lr, =(HighCol+8)
\r
739 mov r0, r0, lsl #26
\r
740 mov r0, r0, lsr #26
\r
743 orr r0, r0, r1, lsl #6
\r
744 orr r0, r0, r0, lsl #8
\r
745 orr r0, r0, r0, lsl #16
\r
756 stmia lr!, {r0-r7} @ 10*8*4
\r
767 ldmfd sp!, {r4-r9,r12}
\r
771 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
774 .global DrawTilesFromCache @ int *hc, int sh, int rlim
\r
776 DrawTilesFromCache:
\r
777 stmfd sp!, {r4-r8,r11,lr}
\r
779 @ cache some stuff to avoid mem access
\r
780 .if OVERRIDE_HIGHCOL
\r
788 ldr lr, =(Pico+0x10000) @ lr=Pico.vram
\r
790 mvn r5, #0 @ r5=prevcode=-1
\r
792 orr r8, r8, r2, lsl #1
\r
793 bne .dtfc_check_rendflags
\r
797 ldr r6, [r0], #4 @ read code
\r
798 movs r1, r6, lsr #16 @ r1=dx;
\r
799 ldmeqfd sp!, {r4-r8,r11,pc} @ dx is never zero, this must be a terminator, return
\r
800 bic r4, r1, #0xfe00
\r
801 add r1, r11, r4 @ r1=pdest
\r
803 mov r7, r6, lsl #16
\r
804 cmp r5, r7, lsr #16
\r
805 beq .dtfc_samecode @ if (code==prevcode)
\r
807 mov r5, r7, lsr #16
\r
809 mov r2, r5, lsl #21
\r
810 mov r2, r2, lsr #17 @ r2=addr=(code&0x7ff)<<4;
\r
811 add r2, r2, r6, lsr #25 @ addr+=ty
\r
813 and r3, r5, #0x6000
\r
814 mov r3, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);
\r
816 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
819 rsbs r4, r4, r8, lsr #1
\r
829 beq .dtfc_SingleColor @ tileline singlecolor
\r
834 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
846 orr r4, r4, r4, lsl #8
\r
847 tst r1, #1 @ not aligned?
\r
848 strneb r4, [r1], #1
\r
849 streqh r4, [r1], #2
\r
853 strneb r4, [r1], #1 @ have a remaining unaligned pixel?
\r
858 beq .dtfc_shadow_blank
\r
861 beq .dtfc_SingleColor @ tileline singlecolor
\r
864 bne .dtfc_TileFlipShHP
\r
866 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
867 .dtfc_TileNormShHP:
\r
871 .dtfc_TileFlipShHP:
\r
875 .dtfc_shadow_blank:
\r
880 strneb r4, [r1], #1
\r
882 orr r6, r6, r6, lsl #8
\r
898 add r4, r4, #7 @ 0-6
\r
902 mov r2, r2, ror #16
\r
903 tst r5, #0x0800 @ flipped?
\r
906 mov r2, r2, ror #16
\r
916 @ check if we have detected layer covered with hi-prio tiles:
\r
917 .dtfc_check_rendflags:
\r
918 ldr r1, =rendstatus
\r
920 tst r2, #(PDRAW_PLANE_HI_PRIO|PDRAW_SHHI_DONE)
\r
922 bic r8, r8, #1 @ sh/hi mode off
\r
923 tst r2, #PDRAW_SHHI_DONE
\r
924 bne .dtfc_loop @ already processed
\r
925 orr r2, r2, #PDRAW_SHHI_DONE
\r
931 orr r6, r6, r6, lsl #8
\r
932 orr r6, r6, r6, lsl #16
\r
934 ldmia r1, {r2,r4,r5,r7}
\r
940 stmia r1!,{r2,r4,r5,r7}
\r
941 bne .dtfc_loop_shprep
\r
943 mvn r5, #0 @ r5=prevcode=-1
\r
948 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
951 .global DrawSpritesSHi @ unsigned char *sprited
\r
959 stmfd sp!, {r4-r11,lr}
\r
960 strb r12,[r0,#2] @ set end marker
\r
961 add r10,r0, #3 @ r10=HighLnSpr end
\r
962 add r10,r10,r3 @ r10=HighLnSpr end
\r
964 .if OVERRIDE_HIGHCOL
\r
972 ldr lr, =(Pico+0x10000) @ lr=Pico.vram
\r
977 ldrb r0, [r10,#-1]!
\r
978 ldr r1, =HighPreSpr
\r
979 @ ldr r8, [sp, #-4]
\r
981 ldmeqfd sp!, {r4-r11,pc} @ end of list
\r
983 add r0, r1, r0, lsl #3
\r
985 ldr r9, [r0, #4] @ sprite[1]
\r
986 mov r2, r9, asr #16 @ r2=sx
\r
988 mov r9, r9, lsl #16
\r
989 mov r3, r9, lsr #31 @ priority
\r
990 mov r9, r9, lsr #16
\r
991 @ orr r9, r9, r8, lsl #31 @ r9=code|sh[31] @@ sh is always on here now
\r
992 and r4, r9, #0x6000
\r
993 orr r9, r9, r4, lsl #16
\r
994 orr r9, r9, #0x90000000 @ r9=scc1 ???? ... <code> (s=shadow/hilight, cc=pal)
\r
995 cmp r12,r9, lsr #28 @ sh/hi with pal3?
\r
996 cmpne r3, #1 @ if not, is it hi prio?
\r
997 bne DrawSpriteSHi @ non-operator low sprite, already drawn
\r
999 ldr r3, [r0] @ sprite[0]
\r
1000 ldr r7, =DrawScanline
\r
1001 mov r6, r3, lsr #28
\r
1002 sub r6, r6, #1 @ r6=width-1 (inc later)
\r
1003 mov r5, r3, lsr #24
\r
1004 and r5, r5, #7 @ r5=height
\r
1006 mov r0, r3, lsl #16 @ r4=sy<<16 (tmp)
\r
1009 sub r7, r7, r0, asr #16 @ r7=row=DrawScanline-sy
\r
1012 movne r0, r5, lsl #3
\r
1014 subne r7, r0, r7 @ if (code&0x1000) row=(height<<3)-1-row; // Flip Y
\r
1016 add r8, r9, r7, lsr #3 @ tile+=row>>3; // Tile number increases going down
\r
1018 mlane r8, r5, r6, r8 @ if (code&0x0800) { tile+=delta*(width-1);
\r
1019 rsbne r5, r5, #0 @ delta=-delta; } // r5=delta now
\r
1021 mov r8, r8, lsl #21
\r
1022 mov r8, r8, lsr #17
\r
1024 add r8, r8, r7, lsl #1 @ tile+=(row&7)<<1; // Tile address
\r
1026 mov r5, r5, lsl #4 @ delta<<=4; // Delta of address
\r
1027 mov r3, r4, lsr #9 @ r3=pal=((code>>9)&0x30);
\r
1029 add r6, r6, #1 @ inc now
\r
1030 adds r0, r2, #0 @ mov sx to r0 and set ZV flags
\r
1031 b .dsprShi_loop_enter
\r
1034 subs r6, r6, #1 @ width--
\r
1036 adds r0, r0, #8 @ sx+=8
\r
1037 add r8, r8, r5 @ tile+=delta
\r
1039 .dsprShi_loop_enter:
\r
1040 ble .dsprShi_loop @ sx <= 0
\r
1044 mov r8, r8, lsl #17
\r
1045 mov r8, r8, lsr #17 @ tile&=0x7fff; // Clip tile address
\r
1047 ldr r2, [lr, r8, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
1048 add r1, r11, r0 @ r1=pdest
\r
1052 cmp r12, r9, lsr #28
\r
1053 beq .dsprShi_shadow
\r
1055 cmp r2, r2, ror #4
\r
1056 beq .dsprShi_SingleColor @ tileline singlecolor
\r
1059 bne .dsprShi_TileFlip
\r
1061 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
1063 .dsprShi_TileNorm:
\r
1067 .dsprShi_TileFlip:
\r
1071 .dsprShi_SingleColor:
\r
1074 orr r4, r4, r4, lsl #8
\r
1075 tst r0, #1 @ not aligned?
\r
1076 strneb r4, [r1], #1
\r
1077 streqh r4, [r1], #2
\r
1081 strneb r4, [r1], #1
\r
1086 beq .dsprShi_shadow_lowpri
\r
1088 cmp r2, r2, ror #4
\r
1089 beq .dsprShi_singlec_sh
\r
1092 bne .dsprShi_TileFlip_sh
\r
1094 @ (r1=pdest, r2=pixels8, r3=pal) r4, r7: scratch, r12: helper pattern
\r
1095 .dsprShi_TileNorm_sh:
\r
1099 .dsprShi_TileFlip_sh:
\r
1103 .dsprShi_singlec_sh:
\r
1104 cmp r2, #0xe0000000
\r
1105 bcc .dsprShi_SingleColor @ normal singlecolor tileline (carry inverted in ARM)
\r
1106 tst r2, #0x10000000
\r
1107 bne .dsprShi_sh_sh
\r
1115 .dsprShi_shadow_lowpri:
\r
1117 bne .dsprShi_TileFlip_sh_lp
\r
1119 .dsprShi_TileNorm_sh_lp:
\r
1120 TileNormSh_onlyop_lp
\r
1123 .dsprShi_TileFlip_sh_lp:
\r
1124 TileFlipSh_onlyop_lp
\r
1129 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1131 .global DrawAllSprites @ unsigned char *sprited, int prio, int sh
\r
1134 ldr r3, =rendstatus
\r
1135 orr r1, r2, r1, lsl #1
\r
1137 tst r12,#(PDRAW_DIRTY_SPRITES|PDRAW_SPRITES_MOVED)
\r
1139 stmfd sp!, {r0,r1,lr}
\r
1140 and r0, r12,#PDRAW_DIRTY_SPRITES
\r
1141 bic r12,r12,#(PDRAW_DIRTY_SPRITES|PDRAW_SPRITES_MOVED)
\r
1144 ldmfd sp!, {r0,r1,lr}
\r
1148 ands r3, r3, #0x7f
\r
1151 @ time to do some real work
\r
1152 stmfd sp!, {r4-r11,lr}
\r
1154 strb r12,[r0,#2] @ set end marker
\r
1156 add r10,r10,r3 @ r10=HighLnSpr end
\r
1158 str r1, [sp, #-4] @ no calls after this point
\r
1160 .if OVERRIDE_HIGHCOL
\r
1168 ldr lr, =(Pico+0x10000) @ lr=Pico.vram
\r
1170 @ + 0 : hhhhvvvv ----hhvv yyyyyyyy yyyyyyyy // v, h: horiz. size
\r
1171 @ + 4 : xxxxxxxx xxxxxxxx pccvhnnn nnnnnnnn // x: x coord + 8
\r
1173 DrawSprite: @ was: unsigned int *sprite, int sh, int acc_sprites
\r
1174 @ draw next sprite
\r
1175 ldrb r0, [r10,#-1]!
\r
1176 ldr r1, =HighPreSpr
\r
1178 mov r2, r0, lsr #7
\r
1180 ldmeqfd sp!, {r4-r11,pc} @ end of list
\r
1181 cmp r2, r8, lsr #1
\r
1182 bne DrawSprite @ wrong priority
\r
1184 add r0, r1, r0, lsl #3
\r
1186 @ stmfd sp!, {r4-r9,r11,lr}
\r
1187 @ orr r8, r2, r1, lsl #4
\r
1189 ldr r3, [r0] @ sprite[0]
\r
1190 ldr r7, =DrawScanline
\r
1191 mov r6, r3, lsr #28
\r
1192 sub r6, r6, #1 @ r6=width-1 (inc later)
\r
1193 mov r5, r3, lsr #24
\r
1194 and r5, r5, #7 @ r5=height
\r
1196 mov r4, r3, lsl #16 @ r4=sy<<16 (tmp)
\r
1200 sub r7, r7, r4, asr #16 @ r7=row=DrawScanline-sy
\r
1202 mov r2, r9, asr #16 @ r2=sx
\r
1203 mov r9, r9, lsl #16
\r
1204 mov r9, r9, lsr #16
\r
1205 orr r9, r9, r8, lsl #31 @ r9=code|sh[31]
\r
1208 movne r4, r5, lsl #3
\r
1210 subne r7, r4, r7 @ if (code&0x1000) row=(height<<3)-1-row; // Flip Y
\r
1212 add r8, r9, r7, lsr #3 @ tile+=row>>3; // Tile number increases going down
\r
1214 mlane r8, r5, r6, r8 @ if (code&0x0800) { tile+=delta*(width-1);
\r
1215 rsbne r5, r5, #0 @ delta=-delta; } // r5=delta now
\r
1217 mov r8, r8, lsl #21
\r
1218 mov r8, r8, lsr #17
\r
1220 add r8, r8, r7, lsl #1 @ tile+=(row&7)<<1; // Tile address
\r
1223 @ cache some stuff to avoid mem access
\r
1224 mov r5, r5, lsl #4 @ delta<<=4; // Delta of address
\r
1225 and r4, r9, #0x6000
\r
1226 orr r9, r9, r4, lsl #16
\r
1227 orrs r9, r9, #0x10000000 @ r9=scc1 ???? ... <code> (s=shadow/hilight, cc=pal)
\r
1229 mov r3, r4, lsr #9 @ r3=pal=((code>>9)&0x30);
\r
1230 orrmi r3, r3, #0x40 @ for sh/hi
\r
1232 add r6, r6, #1 @ inc now
\r
1233 adds r0, r2, #0 @ mov sx to r0 and set ZV flags
\r
1234 b .dspr_loop_enter
\r
1237 subs r6, r6, #1 @ width--
\r
1239 adds r0, r0, #8 @ sx+=8
\r
1240 add r8, r8, r5 @ tile+=delta
\r
1243 ble .dspr_loop @ sx <= 0
\r
1247 mov r8, r8, lsl #17
\r
1248 mov r8, r8, lsr #17 @ tile&=0x7fff; // Clip tile address
\r
1250 ldr r2, [lr, r8, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
1251 add r1, r11, r0 @ r1=pdest
\r
1255 cmp r12, r9, lsr #28
\r
1258 cmp r2, r2, ror #4
\r
1259 beq .dspr_SingleColor @ tileline singlecolor
\r
1262 bne .dspr_TileFlip
\r
1264 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
1275 cmp r2, #0xe0000000
\r
1276 bcs .dspr_TileNorm_sh @ op. tileline, markop. XXX: maybe add a spec. handler?
\r
1278 .dspr_SingleColor:
\r
1281 orr r4, r4, r4, lsl #8
\r
1282 tst r0, #1 @ not aligned?
\r
1283 strneb r4, [r1], #1
\r
1284 streqh r4, [r1], #2
\r
1288 strneb r4, [r1], #1
\r
1292 cmp r2, r2, ror #4
\r
1293 beq .dspr_singlec_sh
\r
1296 bne .dspr_TileFlip_sh
\r
1298 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
1299 .dspr_TileNorm_sh:
\r
1303 .dspr_TileFlip_sh:
\r
1308 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1310 .global DrawWindow @ int tstart, int tend, int prio, int sh // int *hcache
\r
1313 stmfd sp!, {r4-r11,lr}
\r
1315 ldr r11, =(Pico+0x22228) @ Pico.video
\r
1316 ldr r10, =DrawScanline
\r
1317 ldrb r12, [r11, #3] @ pvid->reg[3]
\r
1320 ldr r4, [r11, #12]
\r
1321 mov r5, r10, lsr #3
\r
1323 mov r10, r10, lsl #1 @ r10=ty
\r
1325 mov r12, r12, lsl #10
\r
1327 tst r4, #1 @ 40 cell mode?
\r
1328 andne r12, r12, #0xf000 @ 0x3c<<10
\r
1329 andeq r12, r12, #0xf800
\r
1330 addne r12, r12, r5, lsl #7
\r
1331 addeq r12, r12, r5, lsl #6 @ nametab
\r
1332 add r12, r12, r0, lsl #2 @ +starttile
\r
1334 ldr r6, =rendstatus
\r
1335 ldr lr, =(Pico+0x10000) @ lr=Pico.vram
\r
1338 @ fetch the first code now
\r
1339 ldrh r7, [lr, r12]
\r
1341 ands r6, r6, #PDRAW_WND_DIFF_PRIO
\r
1344 eoreq r8, r2, r7, lsr #15 @ do prio bits differ?
\r
1346 ldmeqfd sp!, {r4-r11,pc} @ yes, assume that whole window uses same priority
\r
1348 orr r6, r6, r3, lsl #8 @ shadow mode
\r
1352 @ cache some stuff to avoid mem access
\r
1353 .if OVERRIDE_HIGHCOL
\r
1355 mov r8, r8, lsl #1 @ cells
\r
1357 mvn r9, #0 @ r9=prevcode=-1
\r
1360 ldr r11,=(HighCol+8)
\r
1361 mov r8, r8, lsl #1 @ cells
\r
1362 mvn r9, #0 @ r9=prevcode=-1
\r
1364 add r1, r11, r0, lsl #4 @ r1=pdest
\r
1368 @ r4,r5 are scratch in this loop
\r
1372 add r12, r12, #2 @ halfwords
\r
1373 ldrh r7, [lr, r12] @ r7=code (int, but from unsigned, no sign extend)
\r
1375 beq .dwloop_end @ done
\r
1377 eor r5, r6, r7, lsr #15
\r
1379 orrne r6, r6, #2 @ wrong pri
\r
1383 beq .dw_samecode @ we know stuff about this tile already
\r
1386 mov r9, r7 @ remember code
\r
1388 movs r2, r9, lsl #20 @ if (code&0x1000)
\r
1389 mov r2, r2, lsl #1
\r
1390 add r2, r10, r2, lsr #17 @ r2=addr=(code&0x7ff)<<4; addr+=ty
\r
1391 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
1393 and r3, r9, #0x6000
\r
1394 mov r3, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);
\r
1396 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
1403 beq .dwloop @ tileline blank
\r
1405 cmp r2, r2, ror #4
\r
1406 beq .dw_SingleColor @ tileline singlecolor
\r
1411 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r0: helper pattern
\r
1421 and r4, r0, r2 @ #0x0000000f
\r
1423 orr r4, r4, r4, lsl #8
\r
1424 orr r4, r4, r4, lsl #16
\r
1426 stmia r1!, {r4,r5}
\r
1427 b .dwloop_nor1 @ we incremeted r1 ourselves
\r
1430 tst r6, #1 @ hi pri?
\r
1431 orreq r3, r3, #0x40
\r
1432 beq .dw_shadow_done
\r
1435 orr r5, r5, r5, lsl #8
\r
1436 orr r5, r5, r5, lsl #16
\r
1445 ldr r0, =rendstatus
\r
1447 and r6, r6, #PDRAW_WND_DIFF_PRIO
\r
1451 ldmfd sp!, {r4-r11,r12}
\r
1455 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1458 @ hilights 2 pixels in RGB444/BGR444 format
\r
1459 .macro TileDoShHi2Pixels444 reg
\r
1460 mov \reg, \reg, ror #12
\r
1461 adds \reg, \reg, #0x40000000
\r
1462 orrcs \reg, \reg, #0xf0000000
\r
1463 mov \reg, \reg, ror #28
\r
1464 adds \reg, \reg, #0x40000000
\r
1465 orrcs \reg, \reg, #0xf0000000
\r
1466 mov \reg, \reg, ror #28
\r
1467 adds \reg, \reg, #0x40000000
\r
1468 orrcs \reg, \reg, #0xf0000000
\r
1469 mov \reg, \reg, ror #24
\r
1470 adds \reg, \reg, #0x40000000
\r
1471 orrcs \reg, \reg, #0xf0000000
\r
1472 mov \reg, \reg, ror #28
\r
1473 adds \reg, \reg, #0x40000000
\r
1474 orrcs \reg, \reg, #0xf0000000
\r
1475 mov \reg, \reg, ror #28
\r
1476 adds \reg, \reg, #0x40000000
\r
1477 orrcs \reg, \reg, #0xf0000000
\r
1478 mov \reg, \reg, ror #12
\r
1482 .global FinalizeLineBGR444 @ int sh
\r
1484 FinalizeLineBGR444:
\r
1485 stmfd sp!, {r4-r6,lr}
\r
1487 ldr lr, =(Pico+0x22228) @ Pico.video
\r
1488 ldr r0, =DrawLineDest
\r
1489 ldrb r12, [lr, #12]
\r
1491 sub r3, lr, #0x128 @ r3=Pico.cram
\r
1494 movne r2, #320/4 @ len
\r
1495 bne .fl_no32colBGR444
\r
1500 addeq r0, r0, #32*2
\r
1502 .fl_no32colBGR444:
\r
1504 beq .fl_noshBGR444
\r
1508 ldrb r12, [lr, #-0x1a] @ 0x2220e ~ dirtyPal
\r
1511 beq .fl_noshBGR444
\r
1513 strb r12, [lr, #-0x1a]
\r
1518 ldmia r3!, {r1,r5,r6,r12}
\r
1520 stmia r4!, {r1,r5,r6,r12}
\r
1521 bne .fl_loopcpBGR444
\r
1523 @ shadowed pixels:
\r
1525 orr r12,r12,#0x0700
\r
1526 orr r12,r12,r12,lsl #16
\r
1527 sub r3, r3, #0x40*2
\r
1528 add r5, r4, #0x80*2
\r
1530 .fl_loopcpBGR444_sh:
\r
1531 ldmia r3!, {r1,r6}
\r
1533 and r1, r12, r1, lsr #1
\r
1534 and r6, r12, r6, lsr #1
\r
1535 stmia r4!, {r1,r6}
\r
1536 stmia r5!, {r1,r6}
\r
1537 bne .fl_loopcpBGR444_sh
\r
1539 @ hilighted pixels:
\r
1540 sub r3, r3, #0x40*2
\r
1542 .fl_loopcpBGR444_hi:
\r
1544 TileDoShHi2Pixels444 r1
\r
1547 bne .fl_loopcpBGR444_hi
\r
1549 sub r3, r4, #0x40*3*2
\r
1554 ldr r12,=rendstatus
\r
1555 eors r6, r6, #1 @ sh is 0
\r
1558 tstne r12,#PDRAW_ACC_SPRITES
\r
1560 .if OVERRIDE_HIGHCOL
\r
1564 mov lr, lr, lsl #1
\r
1567 ldr r1, =(HighCol+8)
\r
1569 mov lr, lr, lsl #1
\r
1576 and r4, lr, r12, lsl #1
\r
1578 and r5, lr, r12, lsr #7
\r
1580 and r6, lr, r12, lsr #15
\r
1582 and r12,lr, r12, lsr #23
\r
1583 ldrh r12,[r3, r12] @ 1c.i.
\r
1584 orr r4, r4, r5, lsl #16
\r
1585 orr r5, r6, r12,lsl #16
\r
1587 stmia r0!, {r4,r5}
\r
1588 bne .fl_loopBGR444
\r
1591 ldmfd sp!, {r4-r6,lr}
\r
1595 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1598 @ Convert 0000bbb0 ggg0rrr0
\r
1599 @ to rrrrrggg gggbbbbb
\r
1601 @ r2,r3 - scratch, lr = 0x001c001c, r8 = 0x08610861
\r
1602 .macro convRGB565 reg
\r
1603 and r2, lr, \reg,lsr #7 @ b
\r
1604 and r3, lr, \reg,lsr #3 @ g
\r
1605 and \reg, lr, \reg,lsl #1 @ r
\r
1606 orr r2, r2, r3, lsl #6
\r
1607 orr \reg, r2, \reg,lsl #11
\r
1609 and r2, r8, \reg,lsr #4
\r
1610 orr \reg, \reg, r2
\r
1613 @ trashes: r2-r8,r12,lr; r8 = 0x08610861; r0,r1 are advanced
\r
1614 .macro vidConvCpyRGB565_local
\r
1615 mov r12, r2, lsr #3 @ repeats
\r
1616 mov lr, #0x001c0000
\r
1617 orr lr, lr, #0x01c @ lr == pattern 0x001c001c
\r
1620 ldmia r1!, {r4-r7}
\r
1635 .global vidConvCpyRGB565
\r
1637 vidConvCpyRGB565: @ void *to, void *from, int pixels
\r
1638 stmfd sp!, {r4-r9,lr}
\r
1640 orr r8, r8, #0x0800
\r
1641 orr r8, r8, r8, lsl #16
\r
1642 vidConvCpyRGB565_local
\r
1643 ldmfd sp!, {r4-r9,lr}
\r
1647 .global PicoDoHighPal555 @ int sh
\r
1650 stmfd sp!, {r4-r9,lr}
\r
1652 ldr r8, =(Pico+0x22228) @ Pico.video
\r
1654 PicoDoHighPal555_nopush:
\r
1655 orr r9, r1, r0, lsl #31 @ 0:called from FinalizeLine555, 31: s/h
\r
1660 strb r1, [r8, #-0x1a] @ 0x2220e ~ dirtyPal
\r
1662 sub r1, r8, #0x128 @ r1=Pico.cram
\r
1665 orr r8, r8, #0x0800
\r
1666 orr r8, r8, r8, lsl #16
\r
1668 vidConvCpyRGB565_local
\r
1671 beq PicoDoHighPal555_end
\r
1675 @ shadowed pixels:
\r
1677 add r4, r3, #0x40*2
\r
1678 orr r12,r12,#0x7300
\r
1679 add r5, r3, #0xc0*2
\r
1680 orr r12,r12,r12,lsl #16
\r
1682 .fl_loopcpRGB555_sh:
\r
1683 ldmia r3!, {r1,r6}
\r
1685 and r1, r12, r1, lsr #1
\r
1686 and r6, r12, r6, lsr #1
\r
1687 stmia r4!, {r1,r6}
\r
1688 stmia r5!, {r1,r6}
\r
1689 bne .fl_loopcpRGB555_sh
\r
1691 @ hilighted pixels:
\r
1692 @ t = ((dpal[i] >> 1) & 0x738e738e) + 0x738e738e;
\r
1693 @ t |= (t >> 4) & 0x08610861;
\r
1695 sub r3, r3, #0x40*2
\r
1697 .fl_loopcpRGB555_hi:
\r
1698 ldmia r3!, {r1,r6}
\r
1699 and r1, r12, r1, lsr #1
\r
1700 and r6, r12, r6, lsr #1
\r
1703 and r5, r8, r1, lsr #4
\r
1704 and r7, r8, r6, lsr #4
\r
1707 stmia r4!, {r1,r6}
\r
1709 bne .fl_loopcpRGB555_hi
\r
1712 PicoDoHighPal555_end:
\r
1714 ldmeqfd sp!, {r4-r9,pc}
\r
1716 ldr r8, =(Pico+0x22228) @ Pico.video
\r
1717 b FinalizeLineRGB555_pal_done
\r
1720 .global FinalizeLine555 @ int sh
\r
1723 stmfd sp!, {r4-r9,lr}
\r
1724 ldr r8, =(Pico+0x22228) @ Pico.video
\r
1726 ldrb r2, [r8, #-0x1a] @ 0x2220e ~ dirtyPal
\r
1729 bne PicoDoHighPal555_nopush
\r
1731 FinalizeLineRGB555_pal_done:
\r
1734 ldr r12,=rendstatus
\r
1735 eors r0, r0, #1 @ sh is 0
\r
1738 tstne r12,#PDRAW_ACC_SPRITES
\r
1741 .if OVERRIDE_HIGHCOL
\r
1743 ldr r0, =DrawLineDest
\r
1748 ldr r0, =DrawLineDest
\r
1749 ldr r1, =(HighCol+8)
\r
1753 ldrb r12, [r8, #12]
\r
1754 mov lr, lr, lsl #1
\r
1757 movne r2, #320/8 @ len
\r
1758 bne .fl_no32colRGB555
\r
1763 bne .fl_32scale_RGB555
\r
1765 addeq r0, r0, #32*2
\r
1767 .fl_no32colRGB555:
\r
1769 #ifdef UNALIGNED_DRAWLINEDEST
\r
1770 @ this is basically for Gizmondo, which has unaligned odd lines in the framebuffer
\r
1779 and r4, lr, r12, lsl #1
\r
1781 and r5, lr, r12, lsr #7
\r
1783 and r6, lr, r12, lsr #15
\r
1785 orr r4, r4, r5, lsl #16
\r
1787 and r5, lr, r12, lsr #23
\r
1789 and r8, lr, r7, lsl #1
\r
1791 orr r5, r6, r5, lsl #16
\r
1793 and r6, lr, r7, lsr #7
\r
1795 and r12,lr, r7, lsr #15
\r
1796 ldrh r12,[r3, r12]
\r
1797 and r7, lr, r7, lsr #23
\r
1799 orr r8, r8, r6, lsl #16
\r
1802 orr r12,r12, r7, lsl #16
\r
1804 stmia r0!, {r4,r5,r8,r12}
\r
1805 bne .fl_loopRGB555
\r
1807 ldmfd sp!, {r4-r9,lr}
\r
1811 .fl_32scale_RGB555:
\r
1813 mov r9, #0x3900 @ f800 07e0 001f | e000 0780 001c | 3800 01e0 0007
\r
1814 orr r9, r9, #0x00e7
\r
1816 #ifdef UNALIGNED_DRAWLINEDEST
\r
1818 bne .fl_32scale_RGB555u
\r
1821 .fl_loop32scale_RGB555:
\r
1825 and r4, lr, r12,lsl #1
\r
1827 and r5, lr, r12,lsr #7
\r
1829 and r4, r4, r9, lsl #2
\r
1830 orr r4, r4, r4, lsl #14 @ r4[31:16] = 1/4 pix_s 0
\r
1831 and r5, r5, r9, lsl #2
\r
1832 sub r6, r5, r5, lsr #2 @ r6 = 3/4 pix_s 1
\r
1833 add r4, r4, r6, lsl #16 @ pix_d 0, 1
\r
1834 and r6, lr, r12,lsr #15
\r
1836 and r12,lr, r12,lsr #23
\r
1837 ldrh r12,[r3, r12]
\r
1838 and r6, r6, r9, lsl #2
\r
1840 mov r5, r5, lsr #1
\r
1841 sub r6, r6, r6, lsr #2 @ r6 = 3/4 pix_s 2
\r
1842 orr r5, r5, r6, lsl #16
\r
1844 and r6, lr, r7, lsl #1
\r
1846 and r12,r12,r9, lsl #2
\r
1847 add r5, r5, r12,lsl #14 @ pix_d 2, 3
\r
1848 and r6, r6, r9, lsl #2
\r
1849 orr r6, r12,r6, lsl #16 @ pix_d 4, 5
\r
1851 and r12,lr, r7, lsr #7
\r
1852 ldrh r12,[r3, r12]
\r
1853 and r10,lr, r7, lsr #15
\r
1854 ldrh r10,[r3, r10]
\r
1855 and r12,r12,r9, lsl #2
\r
1856 sub r8, r12,r12,lsr #2 @ r8 = 3/4 pix_s 1
\r
1857 add r8, r8, r6, lsr #18
\r
1858 and r7, lr, r7, lsr #23
\r
1860 and r10,r10,r9, lsl #2
\r
1861 orr r8, r8, r10,lsl #15
\r
1862 add r8, r8, r12,lsl #15 @ pix_d 6, 7
\r
1863 sub r10,r10,r10,lsr #2 @ r10= 3/4 pix_s 2
\r
1864 and r7, r7, r9, lsl #2
\r
1865 add r10,r10,r7, lsr #2 @ += 1/4 pix_s 3
\r
1866 orr r10,r10,r7, lsl #16 @ pix_d 8, 9
\r
1870 stmia r0!, {r4,r5,r6,r8,r10}
\r
1871 bne .fl_loop32scale_RGB555
\r
1874 ldmfd sp!, {r4-r9,lr}
\r
1877 #ifdef UNALIGNED_DRAWLINEDEST
\r
1878 @ unaligned versions of loops
\r
1879 @ warning: starts drawing 2bytes before dst
\r
1882 sub r0, r0, #2 @ initial adjustment
\r
1889 and r6, lr, r12,lsl #1
\r
1891 and r5, lr, r12,lsr #7
\r
1893 orr r4, r8, r6, lsl #16
\r
1895 and r6, lr, r12,lsr #15
\r
1897 and r8, lr, r12,lsr #23
\r
1899 orr r5, r5, r6, lsl #16
\r
1901 and r6, lr, r7, lsl #1
\r
1903 and r12,lr, r7, lsr #7
\r
1904 ldrh r12,[r3, r12]
\r
1905 orr r6, r8, r6, lsl #16
\r
1907 and r8, lr, r7, lsr #15
\r
1909 and r7, lr, r7, lsr #23
\r
1912 orr r12,r12,r8, lsl #16
\r
1915 stmia r0!, {r4,r5,r6,r12}
\r
1916 bne .fl_loopRGB555u
\r
1920 ldmfd sp!, {r4-r9,lr}
\r
1924 .fl_32scale_RGB555u:
\r
1925 sub r0, r0, #2 @ initial adjustment
\r
1928 @ r9 f800 07e0 001f | e000 0780 001c | 3800 01e0 0007
\r
1929 .fl_loop32scale_RGB555u:
\r
1933 and r6, lr, r12,lsl #1
\r
1935 and r5, lr, r12,lsr #7
\r
1937 and r6, r6, r9, lsl #2
\r
1938 orr r4, r4, r6, lsl #16 @ r4 = pix_d -1, 0
\r
1940 and r5, r5, r9, lsl #2
\r
1941 sub r8, r5, r5, lsr #2 @ r8 = 3/4 pix_s 1
\r
1942 add r6, r8, r6, lsr #2 @ r6 = (1/4 pix_s 0) + (3/4 pix_s 1)
\r
1943 orr r5, r6, r5, lsl #15
\r
1945 and r6, lr, r12,lsr #15
\r
1947 and r12,lr, r12,lsr #23
\r
1948 ldrh r12,[r3, r12]
\r
1949 and r6, r6, r9, lsl #2
\r
1950 add r5, r5, r6, lsl #15 @ r5 = pix_d 1, 2
\r
1952 and r8, lr, r7, lsl #1
\r
1954 and r10,lr, r7, lsr #7
\r
1955 ldrh r10,[r3, r10]
\r
1956 and r12,r12,r9, lsl #2
\r
1957 sub r6, r6, r6, lsr #2 @ r6 = 3/4 pix_s 2
\r
1958 add r6, r6, r12,lsr #2
\r
1959 orr r6, r6, r12,lsl #16 @ r6 = pix_d 3, 4
\r
1961 and r8, r8, r9, lsl #2
\r
1962 and r10,r10,r9, lsl #2
\r
1963 sub r12,r10,r10,lsr #2 @ r12 = 3/4 pix_s 5
\r
1964 orr r8, r8, r8, lsl #14
\r
1965 add r8, r8, r12,lsl #16 @ r8 = pix_d 5, 6
\r
1966 and r12,lr, r7, lsr #15
\r
1967 ldrh r12,[r3, r12]
\r
1968 and r7, lr, r7, lsr #23
\r
1970 and r12,r12,r9, lsl #2
\r
1972 mov r10,r10, lsr #1
\r
1973 sub r12,r12,r12,lsr #2 @ r12 = 3/4 pix_s 6
\r
1974 orr r10,r10,r12,lsl #16
\r
1975 and r7, r7, r9, lsl #2
\r
1976 add r10,r10,r7, lsl #14 @ r10 = pix_d 7, 8
\r
1980 stmia r0!, {r4,r5,r6,r8,r10}
\r
1982 bne .fl_loop32scale_RGB555u
\r
1987 ldmfd sp!, {r4-r9,lr}
\r
1990 #endif /* UNALIGNED_DRAWLINEDEST */
\r
1993 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1996 .global blockcpy @ void *dst, void *src, size_t n
\r
1999 stmfd sp!, {r4,r5}
\r
2000 mov r2, r2, lsr #4
\r
2002 ldmia r1!, {r3-r5,r12}
\r
2004 stmia r0!, {r3-r5,r12}
\r
2006 ldmfd sp!, {r4,r5}
\r
2010 .global blockcpy_or @ void *dst, void *src, size_t n, int pat
\r
2013 stmfd sp!, {r4-r6}
\r
2014 orr r3, r3, r3, lsl #8
\r
2015 orr r3, r3, r3, lsl #16
\r
2016 mov r2, r2, lsr #4
\r
2018 ldmia r1!, {r4-r6,r12}
\r
2024 stmia r0!, {r4-r6,r12}
\r
2025 bne blockcpy_loop_or
\r
2026 ldmfd sp!, {r4-r6}
\r
2029 @ vim:filetype=armasm
\r