1 @ assembly "optimized" version of some funtions from draw.c
\r
2 @ this is highly specialized, be careful if changing related C code!
\r
4 @ (c) Copyright 2006, notaz
\r
5 @ All Rights Reserved
\r
14 .extern DrawLineDest
\r
15 .extern DrawStripVSRam
\r
16 .extern DrawStripInterlace
\r
20 .macro TilePixel pat lsrr offs
\r
24 ands r4, \pat, r2, lsr #\lsrr
\r
27 strneb r4, [r1,#\offs]
\r
30 @ TileNorm (r1=pdest, r2=pixels8, r3=pal) r4: scratch, pat: register with helper pattern 0xf
\r
32 TilePixel \pat, 12, 0 @ #0x0000f000
\r
33 TilePixel \pat, 8, 1 @ #0x00000f00
\r
34 TilePixel \pat, 4, 2 @ #0x000000f0
\r
35 TilePixel \pat, 0, 3 @ #0x0000000f
\r
36 TilePixel \pat, 28, 4 @ #0xf0000000
\r
37 TilePixel \pat, 24, 5 @ #0x0f000000
\r
38 TilePixel \pat, 20, 6 @ #0x00f00000
\r
39 TilePixel \pat, 16, 7 @ #0x000f0000
\r
42 @ TileFlip (r1=pdest, r2=pixels8, r3=pal) r4: scratch, pat: register with helper pattern 0xf
\r
44 TilePixel \pat, 16, 0 @ #0x000f0000
\r
45 TilePixel \pat, 20, 1 @ #0x00f00000
\r
46 TilePixel \pat, 24, 2 @ #0x0f000000
\r
47 TilePixel \pat, 28, 3 @ #0xf0000000
\r
48 TilePixel \pat, 0, 4 @ #0x0000000f
\r
49 TilePixel \pat, 4, 5 @ #0x000000f0
\r
50 TilePixel \pat, 8, 6 @ #0x00000f00
\r
51 TilePixel \pat, 12, 7 @ #0x0000f000
\r
54 @ shadow/hilight mode
\r
56 @ this one is for hi priority layer
\r
57 .macro TilePixelShHP pat lsrr offs
\r
58 TilePixel \pat, \lsrr, \offs
\r
59 ldreqb r4, [r1,#\offs]
\r
62 streqb r4, [r1,#\offs]
\r
65 @ TileNorm (r1=pdest, r2=pixels8, r3=pal) r4: scratch, pat: register with helper pattern 0xf
\r
66 .macro TileNormShHP pat
\r
67 TilePixelShHP \pat, 12, 0 @ #0x0000f000
\r
68 TilePixelShHP \pat, 8, 1 @ #0x00000f00
\r
69 TilePixelShHP \pat, 4, 2 @ #0x000000f0
\r
70 TilePixelShHP \pat, 0, 3 @ #0x0000000f
\r
71 TilePixelShHP \pat, 28, 4 @ #0xf0000000
\r
72 TilePixelShHP \pat, 24, 5 @ #0x0f000000
\r
73 TilePixelShHP \pat, 20, 6 @ #0x00f00000
\r
74 TilePixelShHP \pat, 16, 7 @ #0x000f0000
\r
77 @ TileFlip (r1=pdest, r2=pixels8, r3=pal) r4: scratch, pat: register with helper pattern 0xf
\r
78 .macro TileFlipShHP pat
\r
79 TilePixelShHP \pat, 16, 0 @ #0x000f0000
\r
80 TilePixelShHP \pat, 20, 1 @ #0x00f00000
\r
81 TilePixelShHP \pat, 24, 2 @ #0x0f000000
\r
82 TilePixelShHP \pat, 28, 3 @ #0xf0000000
\r
83 TilePixelShHP \pat, 0, 4 @ #0x0000000f
\r
84 TilePixelShHP \pat, 4, 5 @ #0x000000f0
\r
85 TilePixelShHP \pat, 8, 6 @ #0x00000f00
\r
86 TilePixelShHP \pat, 12, 7 @ #0x0000f000
\r
90 @ TileSingleSh (r1=pdest, r2=pixels8, r3=pal) r4,r7: scratch, r0=sx; r12: helper pattern 0xf
\r
92 tst r0, #1 @ not aligned?
\r
111 strneb r4, [r1], #1
\r
114 @ TileSingleHi (r1=pdest, r2=pixels8, r3=pal) r4,r7: scratch, r0=sx, r12: register with helper pattern 0xf
\r
115 .macro TileSingleHi
\r
116 tst r1, #1 @ not aligned?
\r
121 bic r4, r4, r7, lsr #1
\r
123 strneb r4, [r1], #1
\r
124 streqh r4, [r1], #2
\r
126 bic r4, r4, r7, lsr #1
\r
130 bic r4, r4, r7, lsr #1
\r
134 bic r4, r4, r7, lsr #1
\r
138 bic r4, r4, r7, lsr #1
\r
140 strneb r4, [r1], #1
\r
143 .macro TileDoShGenPixel shift ofs
\r
145 ands r4, r12, r2, lsr #\shift
\r
154 strb r4, [r1,#\ofs]
\r
157 ldrb r4, [r1,#\ofs]
\r
159 strb r4, [r1,#\ofs]
\r
162 ldrb r4, [r1,#\ofs]
\r
165 strb r4, [r1,#\ofs]
\r
169 @ TileFlipSh (r1=pdest, r2=pixels8, r3=pal) r4,r7: scratch, r0=sx, r12: register with helper pattern 0xf
\r
171 TileDoShGenPixel 16, 0 @ #0x000f0000
\r
172 TileDoShGenPixel 20, 1 @ #0x00f00000
\r
173 TileDoShGenPixel 24, 2 @ #0x0f000000
\r
174 TileDoShGenPixel 28, 3 @ #0xf0000000
\r
175 TileDoShGenPixel 0, 4 @ #0x0000000f
\r
176 TileDoShGenPixel 4, 5 @ #0x000000f0
\r
177 TileDoShGenPixel 8, 6 @ #0x00000f00
\r
178 TileDoShGenPixel 12, 7 @ #0x0000f000
\r
181 @ TileNormSh (r1=pdest, r2=pixels8, r3=pal) r4,r7: scratch, r0=sx, r12: register with helper pattern 0xf
\r
183 TileDoShGenPixel 12, 0 @ #0x0000f000
\r
184 TileDoShGenPixel 8, 1 @ #0x00000f00
\r
185 TileDoShGenPixel 4, 2 @ #0x000000f0
\r
186 TileDoShGenPixel 0, 3 @ #0x0000000f
\r
187 TileDoShGenPixel 28, 4 @ #0xf0000000
\r
188 TileDoShGenPixel 24, 5 @ #0x0f000000
\r
189 TileDoShGenPixel 20, 6 @ #0x00f00000
\r
190 TileDoShGenPixel 16, 7 @ #0x000f0000
\r
194 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
198 @ int nametab; // 0x00
\r
199 @ int line; // 0x04
\r
200 @ int hscroll; // 0x08
\r
201 @ int xmask; // 0x0C
\r
202 @ int *hc; // 0x10 (pointer to cache buffer)
\r
203 @ int cells; // 0x14
\r
206 @ int DrawLayer(int plane, int *hcache, int maxcells, int sh)
\r
208 .global DrawLayer @ int plane, int *hcache, int maxcells, int sh
\r
211 stmfd sp!, {r4-r11,lr}
\r
213 ldr r11, =(Pico+0x22228) @ Pico.video
\r
215 mov r6, r1 @ hcache
\r
216 orr r9, r2, r3, lsl #31 @ r9=maxcells|(sh<<31)
\r
218 ldrb r7, [r11, #16] @ ??hh??ww
\r
221 orr r1, r1, #0x00ff
\r
225 biclt r1, r1, #0xfc00
\r
226 biceq r1, r1, #0xfe00
\r
227 bicgt r1, r1, #0xff00 @ r1=ymask=(height<<8)|0xff; ...; // Y Mask in pixels
\r
231 subge r10, r10, #1 @ r10=shift[width] (5,6,6,7)
\r
235 mov r5, r8, lsl r10
\r
236 sub r5, r5, #1 @ r5=xmask
\r
240 ldreqb r12, [r11, #2]
\r
241 moveq r12, r12, lsl #10
\r
242 ldrneb r12, [r11, #4]
\r
243 movne r12, r12, lsl #13
\r
244 and r12, r12, #(7<<13) @ r12=(ts->nametab<<1) (halfword compliant)
\r
248 ldr lr, =(Pico+0x10000) @ lr=Pico.vram
\r
250 ldrh r8, [r11, #12]
\r
251 mov r4, r8, lsr #8 @ pvid->reg[13]
\r
252 mov r4, r4, lsl #10 @ htab=pvid->reg[13]<<9; (halfwords)
\r
253 ldrb r7, [r11, #11]
\r
255 addne r4, r4, r2, lsl #2 @ htab+=Scanline<<1; // Offset by line
\r
257 biceq r4, r4, #0x1f @ htab&=~0xf; // Offset by tile
\r
258 add r4, r4, r0, lsl #1 @ htab+=plane
\r
259 bic r4, r4, #0x00ff0000 @ just in case
\r
260 ldrh r3, [lr, r4] @ r3=hscroll
\r
263 bne .DrawStrip_vsscroll
\r
265 @ Get vertical scroll value:
\r
266 add r7, lr, #0x012000
\r
267 add r7, r7, #0x000180 @ r7=Pico.vsram (Pico+0x22180)
\r
272 bne .DrawStrip_interlace
\r
275 movne r7, r7, lsr #16
\r
277 @ Find the line in the name table
\r
281 add r10, r10, #1 @ shift[width]++
\r
282 add r12, r12, r4, lsl r10 @ nametab+=(ts.line>>3)<<shift[width];
\r
284 @ ldmia r0, {r1,r2,r3,r5,r6,r9} @ r2=line, r3=ts->hscroll, r5=ts->xmask, r6=ts->hc, r9=ts->cells
\r
285 @ mov r12,r1, lsl #1 @ r12=(ts->nametab<<1) (halfword compliant)
\r
288 mov r10,r10, lsl #1 @ r10=ty=(ts->line&7)<<1;
\r
289 orr r10,r10, r9, lsl #24
\r
292 mov r8, r8, lsr #3 @ r8=tilex=(-ts->hscroll)>>3
\r
296 add r7, r1, #1 @ r7=dx=((ts->hscroll-1)&7)+1
\r
300 orrne r10,r10, #1<<23 @ r10=(cells<<24|sh<<23|hi_not_empty<<22|ty)
\r
301 movne r3, #0x40 @ default to shadowed pal on sh mode
\r
303 mvn r9, #0 @ r9=prevcode=-1
\r
306 subeq r10,r10, #0x01000000 @ we will loop cells+1 times, so loop less when there is no scroll
\r
308 @ cache some stuff to avoid mem access
\r
310 add r1, r11, r7 @ r1=pdest
\r
314 @ r4 & r7 are scratch in this loop
\r
315 .dsloop: @ 40-41 times
\r
318 subs r10,r10, #0x01000000
\r
324 add r7, lr, r7, lsl #1 @ Pico.vram+((tilex&ts->xmask) as halfwords)
\r
325 ldrh r7, [r7, r12] @ r7=code (int, but from unsigned, no sign extend)
\r
328 bne .DrawStrip_hiprio
\r
331 beq .DrawStrip_samecode @ we know stuff about this tile already
\r
333 mov r9, r7 @ remember code
\r
335 movs r2, r9, lsl #20 @ if (code&0x1000)
\r
337 @ bic r7, r10,#0xff000000
\r
338 @ add r2, r7, r2, lsr #17 @ r2=addr=(code&0x7ff)<<4; addr+=ty
\r
339 add r2, r2, r10, lsl #17
\r
340 mov r2, r2, lsr #17
\r
341 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
344 and r3, r9, #0x6000
\r
345 add r3, r7, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);
\r
347 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
349 .DrawStrip_samecode:
\r
351 beq .dsloop @ tileline blank
\r
354 beq .DrawStrip_SingleColor @ tileline singlecolor
\r
357 beq .DrawStrip_TileNorm
\r
359 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r0: helper pattern
\r
363 .DrawStrip_TileNorm:
\r
367 .DrawStrip_SingleColor:
\r
370 orr r4, r4, r4, lsl #8
\r
371 tst r1, #1 @ not aligned?
\r
372 strneb r4, [r1], #1
\r
373 streqh r4, [r1], #2
\r
377 strneb r4, [r1], #1 @ have a remaining unaligned pixel?
\r
378 b .dsloop_nor1 @ we incremeted r1 ourselves
\r
381 tst r10, #0x00c00000
\r
382 beq .DrawStrip_hiprio_maybempt
\r
384 orr r7, r7, r0, lsl #16
\r
385 orr r7, r7, r10, lsl #25 @ (ty<<25)
\r
387 eorne r7, r7, #7<<26 @ if(code&0x1000) cval^=7<<26;
\r
388 str r7, [r6], #4 @ cache hi priority tile
\r
392 .DrawStrip_hiprio_maybempt:
\r
394 beq .dsloop @ must've been empty, otherwise we wouldn't get here
\r
395 mov r9, r7 @ remember code
\r
396 movs r2, r9, lsl #20 @ if (code&0x1000)
\r
398 add r2, r2, r10, lsl #17
\r
399 mov r2, r2, lsr #17
\r
400 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
401 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
403 orrne r10, r10, #1<<22
\r
404 bne .DrawStrip_hiprio
\r
409 str r0, [r6] @ terminate the cache list
\r
411 ldmfd sp!, {r4-r11,lr}
\r
415 .DrawStrip_vsscroll:
\r
416 @ shit, we have 2-cell column based vscroll
\r
417 @ let the c code handle this (for now)
\r
419 @ int nametab; // 0x00
\r
420 @ int line; // 0x04
\r
421 @ int hscroll; // 0x08
\r
422 @ int xmask; // 0x0C
\r
423 @ int *hc; // 0x10 (pointer to cache buffer)
\r
424 @ int cells; // 0x14
\r
427 orr r2, r1, r10, lsl #24 @ ts.line=ymask|(shift[width]<<24); // save some stuff instead of line
\r
429 mov r0, r12, lsr #1 @ halfwords
\r
431 stmia sp, {r0,r2,r3,r5,r6,r9}
\r
434 bl DrawStripVSRam @ struct TileStrip *ts, int plane
\r
437 ldmfd sp!, {r4-r11,lr}
\r
440 @ interlace mode 2? Sonic 2?
\r
441 .DrawStrip_interlace:
\r
443 moveq r7, r7, lsl #21
\r
444 movne r7, r7, lsl #5
\r
446 @ Find the line in the name table
\r
447 add r2, r7, r2, lsl #22 @ r2=(vscroll+(Scanline<<1))<<21 (11 bits);
\r
448 orr r1, r1, #0x80000000
\r
449 and r2, r2, r1, ror #10 @ &((ymask<<1)|1)<<21;
\r
450 mov r2, r2, lsr #21
\r
452 mov r12, r12, lsr #1 @ halfwords
\r
453 add r0, r12, r4, lsl r10 @ nametab+=(ts.line>>4)<<shift[width];
\r
457 stmia sp, {r0,r2,r3,r5,r6,r9}
\r
460 bl DrawStripInterlace @ struct TileStrip *ts
\r
463 ldmfd sp!, {r4-r11,lr}
\r
468 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
471 .global BackFill @ int reg7, int sh
\r
474 stmfd sp!, {r4-r9,lr}
\r
476 ldr lr, =(HighCol+8)
\r
478 mov r0, r0, lsl #26
\r
479 mov r0, r0, lsr #26
\r
480 orr r0, r0, r1, lsl #6
\r
481 orr r0, r0, r0, lsl #8
\r
482 orr r0, r0, r0, lsl #16
\r
493 stmia lr!, {r0-r7} @ 10*8*4
\r
504 ldmfd sp!, {r4-r9,r12}
\r
508 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
511 .global DrawTilesFromCache @ int *hc, int sh
\r
513 DrawTilesFromCache:
\r
514 stmfd sp!, {r4-r8,r11,lr}
\r
516 mvn r5, #0 @ r5=prevcode=-1
\r
519 @ cache some stuff to avoid mem access
\r
521 ldr lr, =(Pico+0x10000) @ lr=Pico.vram
\r
526 ldr r6, [r0], #4 @ read code
\r
527 movs r1, r6, lsr #16 @ r1=dx;
\r
528 ldmeqfd sp!, {r4-r8,r11,pc} @ dx is never zero, this must be a terminator, return
\r
529 bic r1, r1, #0xfe00
\r
530 add r1, r11, r1 @ r1=pdest
\r
533 @ bne .dtfc_shadow @ this is a rare case, so we jump when it happens, not when it doesn't
\r
534 @.dtfc_shadow_done:
\r
536 mov r7, r6, lsl #16
\r
537 cmp r5, r7, lsr #16
\r
538 beq .dtfc_samecode @ if (code==prevcode)
\r
540 mov r5, r7, lsr #16
\r
542 mov r2, r5, lsl #21
\r
543 mov r2, r2, lsr #17 @ r2=addr=(code&0x7ff)<<4;
\r
544 add r2, r2, r6, lsr #25 @ addr+=ty
\r
546 and r3, r5, #0x6000
\r
547 mov r3, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);
\r
549 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
559 beq .dtfc_SingleColor @ tileline singlecolor
\r
564 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
575 orr r4, r4, r4, lsl #8
\r
576 tst r1, #1 @ not aligned?
\r
577 strneb r4, [r1], #1
\r
578 streqh r4, [r1], #2
\r
582 strneb r4, [r1], #1 @ have a remaining unaligned pixel?
\r
587 beq .dtfc_shadow_blank
\r
590 beq .dtfc_SingleColor @ tileline singlecolor
\r
593 beq .dtfc_TileNormShHP
\r
595 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
599 .dtfc_TileNormShHP:
\r
603 .dtfc_shadow_blank:
\r
640 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
643 .global DrawSpritesFromCache @ int *hc, int sh
\r
645 DrawSpritesFromCache:
\r
646 stmfd sp!, {r4-r11,lr}
\r
648 @ cache some stuff to avoid mem access
\r
650 ldr lr, =(Pico+0x10000) @ lr=Pico.vram
\r
651 mov r6, r1, lsl #31
\r
658 ldr r9, [r10], #4 @ read code
\r
660 ldmeqfd sp!, {r4-r11,pc}
\r
662 mov r4, r9, lsl #28
\r
664 orr r6, r6, r4, lsr #30
\r
665 add r6, r6, #1 @ r6=s1cc???? ... ?????www (s=shadow/hilight, cc=pal, w=width)
\r
668 add r5, r5, #1 @ r5=delta
\r
670 rsbne r5, r5, #0 @ Flip X
\r
673 mov r2, r9, lsr #17
\r
674 mov r8, r2, lsl #1 @ tile=((unsigned int)code>>17)<<1;
\r
676 and r3, r9, #0x30 @ r3=pal=(code&0x30);
\r
679 orr r6, r6, r3, lsl #24
\r
681 mov r0, r9, lsl #16
\r
682 mov r0, r0, asr #22 @ sx=(code<<16)>>22
\r
683 adds r0, r0, #0 @ set ZV
\r
684 b .dsfc_inloop_enter
\r
694 .dsfc_inloop_enter:
\r
699 mov r8, r8, lsl #17
\r
700 mov r8, r8, lsr #17 @ tile&=0x7fff; // Clip tile address
\r
702 ldr r2, [lr, r8, lsl #1] @ pack=*(unsigned int *)(Pico.vram+tile); // Get 8 pixels
\r
706 add r1, r11, r0 @ r1=pdest
\r
708 cmp r12, r6, lsr #28
\r
712 beq .dsfc_SingleColor @ tileline singlecolor
\r
717 @ TileFlip (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
726 tst r0, #1 @ not aligned?
\r
729 orr r4, r4, r4, lsl #8
\r
730 strneb r4, [r1], #1
\r
731 streqh r4, [r1], #2
\r
735 strneb r4, [r1], #1
\r
740 beq .dsfc_singlec_sh
\r
743 beq .dsfc_TileNorm_sh
\r
745 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
754 cmp r2, #0xe0000000
\r
755 bcc .dsfc_SingleColor @ normal singlecolor tileline (carry inverted in ARM)
\r
756 tst r2, #0x10000000
\r
767 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
769 @ + 0 : hhhhvvvv ab--hhvv yyyyyyyy yyyyyyyy // a: offscreen h, b: offs. v, h: horiz. size
\r
770 @ + 4 : xxxxxxxx xxxxxxxx pccvhnnn nnnnnnnn // x: x coord + 8
\r
772 .global DrawSprite @ unsigned int *sprite, int **hc, int sh
\r
775 stmfd sp!, {r4-r9,r11,lr}
\r
777 ldr r3, [r0] @ sprite[0]
\r
778 mov r6, r3, lsr #28
\r
779 sub r6, r6, #1 @ r6=width-1 (inc later)
\r
780 mov r5, r3, lsr #24
\r
781 and r5, r5, #7 @ r5=height
\r
783 mov r4, r3, lsl #16 @ r4=sy<<16 (tmp)
\r
787 sub r7, r7, r4, asr #16 @ r7=row=Scanline-sy
\r
791 mov r2, r9, asr #16 @ r2=sx
\r
792 bic r9, r9, #0xfe000000
\r
793 orrne r9, r9, #1<<31 @ r9=code|(sh<<31)
\r
796 movne r4, r5, lsl #3
\r
798 subne r7, r4, r7 @ if (code&0x1000) row=(height<<3)-1-row; // Flip Y
\r
800 mov r8, r9, lsl #21
\r
801 mov r8, r8, lsr #21
\r
802 add r8, r8, r7, lsr #3 @ tile+=row>>3; // Tile number increases going down
\r
805 mlane r8, r5, r6, r8 @ if (code&0x0800) { tile+=delta*(width-1);
\r
806 rsbne r5, r5, #0 @ delta=-delta; } // r5=delta now
\r
810 add r8, r8, r7, lsl #1 @ tile+=(row&7)<<1; // Tile address
\r
813 bne .dspr_cache @ if(code&0x8000) // high priority - cache it
\r
815 @ cache some stuff to avoid mem access
\r
817 ldr lr, =(Pico+0x10000) @ lr=Pico.vram
\r
820 mov r5, r5, lsl #4 @ delta<<=4; // Delta of address
\r
821 and r4, r9, #0x6000
\r
822 orr r9, r9, r4, lsl #16
\r
823 orr r9, r9, #0x10000000 @ r9=scc1 ???? ... <code> (s=shadow/hilight, cc=pal)
\r
826 mov r3, r4, lsr #9 @ r3=pal=((code>>9)&0x30);
\r
827 orrne r3, r3, #0x40 @ shadow by default
\r
829 add r6, r6, #1 @ inc now
\r
830 adds r0, r2, #0 @ mov sx to r0 and set ZV flags
\r
834 subs r6, r6, #1 @ width--
\r
835 ldmeqfd sp!, {r4-r9,r11,pc}@ return
\r
836 adds r0, r0, #8 @ sx+=8
\r
837 add r8, r8, r5 @ tile+=delta
\r
840 ble .dspr_loop @ sx <= 0
\r
842 ldmgefd sp!, {r4-r9,r11,pc}@ return
\r
844 mov r8, r8, lsl #17
\r
845 mov r8, r8, lsr #17 @ tile&=0x7fff; // Clip tile address
\r
847 ldr r2, [lr, r8, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
851 add r1, r11, r0 @ r1=pdest
\r
853 cmp r12, r9, lsr #28
\r
857 beq .dspr_SingleColor @ tileline singlecolor
\r
862 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
874 orr r4, r4, r4, lsl #8
\r
875 tst r0, #1 @ not aligned?
\r
876 strneb r4, [r1], #1
\r
877 streqh r4, [r1], #2
\r
881 strneb r4, [r1], #1
\r
886 beq .dspr_singlec_sh
\r
889 beq .dspr_TileNorm_sh
\r
891 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
900 cmp r2, #0xe0000000
\r
901 bcc .dspr_SingleColor @ normal tileline
\r
902 tst r2, #0x10000000
\r
913 @ *(*hc)++ = (tile<<16)|((code&0x0800)<<5)|((sx<<6)&0x0000ffc0)|((code>>9)&0x30)|((sprite[0]>>24)&0xf);
\r
914 mov r4, r8, lsl #16 @ tile
\r
916 orrne r4, r4, #0x10000 @ code&0x0800
\r
917 mov r2, r2, lsl #22
\r
918 orr r4, r4, r2, lsr #16 @ (sx<<6)&0x0000ffc0
\r
919 and r2, r9, #0x6000
\r
920 orr r4, r4, r2, lsr #9 @ (code>>9)&0x30
\r
921 mov r2, r3, lsl #12
\r
922 orr r4, r4, r2, lsr #28 @ (sprite[0]>>24)&0xf
\r
928 ldmfd sp!, {r4-r9,r11,lr}
\r
931 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
933 .global DrawWindow @ int tstart, int tend, int prio, int sh // int *hcache
\r
936 stmfd sp!, {r4-r11,lr}
\r
938 ldr r11, =(Pico+0x22228) @ Pico.video
\r
939 ldrb r12, [r11, #3] @ pvid->reg[3]
\r
940 mov r12, r12, lsl #10
\r
944 mov r5, r10, lsr #3
\r
946 mov r10, r10, lsl #1 @ r10=ty
\r
949 tst r4, #1 @ 40 cell mode?
\r
950 andne r12, r12, #0xf000 @ 0x3c<<10
\r
951 andeq r12, r12, #0xf800
\r
952 addne r12, r12, r5, lsl #7
\r
953 addeq r12, r12, r5, lsl #6 @ nametab
\r
954 add r12, r12, r0, lsl #2 @ +starttile
\r
956 ldr lr, =(Pico+0x10000) @ lr=Pico.vram
\r
958 @ fetch the first code now
\r
961 ldr r6, =rendstatus
\r
963 ands r6, r6, #2 @ we care about bit 1 only
\r
965 bne .dw_no_sameprio
\r
967 cmp r2, r7, lsr #15
\r
968 ldmnefd sp!, {r4-r11,pc} @ assume that whole window uses same priority
\r
971 orr r6, r6, r3, lsl #8 @ shadow mode
\r
974 mov r8, r8, lsl #1 @ cells
\r
976 mvn r9, #0 @ r9=prevcode=-1
\r
978 @ cache some stuff to avoid mem access
\r
979 ldr r11,=(HighCol+8)
\r
980 add r1, r11, r0, lsl #4 @ r1=pdest
\r
984 @ r4,r5 & r7 are scratch in this loop
\r
989 add r12, r12, #2 @ halfwords
\r
990 beq .dwloop_end @ done
\r
992 ldrh r7, [lr, r12] @ r7=code (int, but from unsigned, no sign extend)
\r
994 eor r5, r6, r7, lsr #15
\r
996 orrne r6, r6, #2 @ wrong pri
\r
1000 beq .dw_samecode @ we know stuff about this tile already
\r
1003 mov r9, r7 @ remember code
\r
1005 movs r2, r9, lsl #20 @ if (code&0x1000)
\r
1006 mov r2, r2, lsl #1
\r
1007 add r2, r10, r2, lsr #17 @ r2=addr=(code&0x7ff)<<4; addr+=ty
\r
1008 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
1010 and r3, r9, #0x6000
\r
1011 mov r3, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);
\r
1013 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
1020 beq .dwloop @ tileline blank
\r
1022 cmp r2, r2, ror #4
\r
1023 beq .dw_SingleColor @ tileline singlecolor
\r
1028 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r0: helper pattern
\r
1037 and r4, r0, r2 @ #0x0000000f
\r
1039 orr r4, r4, r4, lsl #8
\r
1040 orr r4, r4, r4, lsl #16
\r
1042 stmia r1!, {r4,r5}
\r
1043 b .dwloop_nor1 @ we incremeted r1 ourselves
\r
1046 tst r6, #1 @ hi pri?
\r
1047 orreq r3, r3, #0x40
\r
1048 beq .dw_shadow_done
\r
1050 tst r4, #0x00000080
\r
1051 biceq r4, r4, #0x000000c0
\r
1052 tst r4, #0x00008000
\r
1053 biceq r4, r4, #0x0000c000
\r
1054 tst r4, #0x00800000
\r
1055 biceq r4, r4, #0x00c00000
\r
1056 tst r4, #0x80000000
\r
1057 biceq r4, r4, #0xc0000000
\r
1060 tst r4, #0x00000080
\r
1061 biceq r4, r4, #0x000000c0
\r
1062 tst r4, #0x00008000
\r
1063 biceq r4, r4, #0x0000c000
\r
1064 tst r4, #0x00800000
\r
1065 biceq r4, r4, #0x00c00000
\r
1066 tst r4, #0x80000000
\r
1067 biceq r4, r4, #0xc0000000
\r
1072 ldr r0, =rendstatus
\r
1078 ldmfd sp!, {r4-r11,r12}
\r
1082 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1085 @ hilights 2 pixels in RGB444/BGR444 format
\r
1086 .macro TileDoShHi2Pixels444 reg
\r
1087 mov \reg, \reg, ror #12
\r
1088 adds \reg, \reg, #0x40000000
\r
1089 orrcs \reg, \reg, #0xf0000000
\r
1090 mov \reg, \reg, ror #28
\r
1091 adds \reg, \reg, #0x40000000
\r
1092 orrcs \reg, \reg, #0xf0000000
\r
1093 mov \reg, \reg, ror #28
\r
1094 adds \reg, \reg, #0x40000000
\r
1095 orrcs \reg, \reg, #0xf0000000
\r
1096 mov \reg, \reg, ror #24
\r
1097 adds \reg, \reg, #0x40000000
\r
1098 orrcs \reg, \reg, #0xf0000000
\r
1099 mov \reg, \reg, ror #28
\r
1100 adds \reg, \reg, #0x40000000
\r
1101 orrcs \reg, \reg, #0xf0000000
\r
1102 mov \reg, \reg, ror #28
\r
1103 adds \reg, \reg, #0x40000000
\r
1104 orrcs \reg, \reg, #0xf0000000
\r
1105 mov \reg, \reg, ror #12
\r
1109 .global FinalizeLineBGR444 @ int sh
\r
1111 FinalizeLineBGR444:
\r
1112 stmfd sp!, {r4-r6,lr}
\r
1114 ldr r0, =DrawLineDest
\r
1116 ldr lr, =(Pico+0x22228) @ Pico.video
\r
1117 sub r3, lr, #0x128 @ r3=Pico.cram
\r
1119 ldrb r12, [lr, #12]
\r
1121 movne r2, #320/4 @ len
\r
1123 addeq r0, r0, #32*2
\r
1124 ldreq r4, =PicoOpt
\r
1127 addeq r0, r0, #32*2
\r
1130 beq .fl_noshBGR444
\r
1134 ldrb r12, [lr, #-0x1a] @ 0x2220e ~ dirtyPal
\r
1137 beq .fl_noshBGR444
\r
1139 strb r12, [lr, #-0x1a]
\r
1145 ldmia r3!, {r1,r5,r6,r12}
\r
1146 stmia r4!, {r1,r5,r6,r12}
\r
1147 bne .fl_loopcpBGR444
\r
1149 @ shadowed pixels:
\r
1151 orr r12,r12,#0x0700
\r
1152 orr r12,r12,r12,lsl #16
\r
1153 sub r3, r3, #0x40*2
\r
1154 add r5, r4, #0x80*2
\r
1156 .fl_loopcpBGR444_sh:
\r
1158 ldmia r3!, {r1,r6}
\r
1159 and r1, r12, r1, lsr #1
\r
1160 and r6, r12, r6, lsr #1
\r
1161 stmia r4!, {r1,r6}
\r
1162 stmia r5!, {r1,r6}
\r
1163 bne .fl_loopcpBGR444_sh
\r
1165 @ hilighted pixels:
\r
1166 sub r3, r3, #0x40*2
\r
1168 .fl_loopcpBGR444_hi:
\r
1170 TileDoShHi2Pixels444 r1
\r
1173 bne .fl_loopcpBGR444_hi
\r
1175 sub r3, r4, #0x40*3*2
\r
1179 ldr r1, =(HighCol+8)
\r
1181 mov lr, lr, lsl #1
\r
1188 and r4, lr, r12, lsl #1
\r
1190 and r5, lr, r12, lsr #7
\r
1192 orr r4, r4, r5, lsl #16
\r
1194 and r5, lr, r12, lsr #15
\r
1196 and r6, lr, r12, lsr #23
\r
1198 orr r5, r5, r6, lsl #16
\r
1200 stmia r0!, {r4,r5}
\r
1201 bne .fl_loopBGR444
\r
1204 ldmfd sp!, {r4-r6,lr}
\r
1208 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1211 @ hilights 2 pixels in RGB555/BGR555 format
\r
1212 .macro TileDoShHi2Pixels555 reg
\r
1213 adds \reg, \reg, #0x40000000
\r
1214 orrcs \reg, \reg, #0xf0000000
\r
1215 mov \reg, \reg, ror #27
\r
1216 adds \reg, \reg, #0x40000000
\r
1217 orrcs \reg, \reg, #0xf0000000
\r
1218 mov \reg, \reg, ror #26
\r
1219 adds \reg, \reg, #0x40000000
\r
1220 orrcs \reg, \reg, #0xf0000000
\r
1221 mov \reg, \reg, ror #27
\r
1222 adds \reg, \reg, #0x40000000
\r
1223 orrcs \reg, \reg, #0xf0000000
\r
1224 mov \reg, \reg, ror #27
\r
1225 adds \reg, \reg, #0x40000000
\r
1226 orrcs \reg, \reg, #0xf0000000
\r
1227 mov \reg, \reg, ror #26
\r
1228 adds \reg, \reg, #0x40000000
\r
1229 orrcs \reg, \reg, #0xf0000000
\r
1230 mov \reg, \reg, ror #27
\r
1234 @ Convert 0000bbb0 ggg0rrr0
\r
1235 @ to rrrrrggg gggbbbbb
\r
1237 @ r2,r3,r9 - scratch, lr = 0x001c001c, r8 = 0x00030003
\r
1238 .macro convRGB565 reg
\r
1239 and r2, lr, \reg,lsl #1
\r
1240 and r9, r8, \reg,lsr #2
\r
1241 orr r2, r2, r9 @ r2=red
\r
1242 and r3, lr, \reg,lsr #7
\r
1243 and r9, r8, \reg,lsr #10
\r
1244 orr r3, r3, r9 @ r3=blue
\r
1245 and \reg, \reg, lr, lsl #3
\r
1246 orr \reg, \reg, \reg,lsl #3 @ green
\r
1247 orr \reg, \reg, r2, lsl #11 @ add red back
\r
1248 orr \reg, \reg, r3 @ add blue back
\r
1251 vidConvCpyRGB565: @ void *to, void *from, int pixels
\r
1252 stmfd sp!, {r4-r9,lr}
\r
1254 mov r12, r2, lsr #3 @ repeats
\r
1255 mov lr, #0x001c0000
\r
1256 orr lr, lr, #0x01c @ lr == pattern 0x001c001c
\r
1257 mov r8, #0x00030000
\r
1258 orr r8, r8, #0x003 @ lr == pattern 0x001c001c
\r
1263 ldmia r1!, {r4-r7}
\r
1275 ldmfd sp!, {r4-r9,lr}
\r
1280 .global FinalizeLineRGB555 @ int sh
\r
1282 FinalizeLineRGB555:
\r
1283 stmfd sp!, {r4-r8,lr}
\r
1284 ldr r5, =(Pico+0x22228) @ Pico.video
\r
1288 ldrb r7, [r5, #-0x1a] @ 0x2220e ~ dirtyPal
\r
1290 beq .fl_noconvRGB555
\r
1292 strb r1, [r5, #-0x1a]
\r
1293 sub r1, r5, #0x128 @ r1=Pico.cram
\r
1296 bl vidConvCpyRGB565
\r
1299 ldr r0, =DrawLineDest
\r
1302 ldrb r12, [r5, #12]
\r
1304 movne r2, #320/8 @ len
\r
1306 ldreq r3, =PicoOpt
\r
1309 addeq r0, r0, #32*2
\r
1313 beq .fl_noshRGB555
\r
1315 beq .fl_noshRGB555
\r
1317 @ shadowed pixels:
\r
1319 orr r12,r12,#0x7300
\r
1320 orr r12,r12,r12,lsl #16
\r
1321 add r4, r3, #0x40*2
\r
1322 add r5, r3, #0xc0*2
\r
1324 .fl_loopcpRGB555_sh:
\r
1326 ldmia r3!, {r1,r6}
\r
1327 and r1, r12, r1, lsr #1
\r
1328 and r6, r12, r6, lsr #1
\r
1329 stmia r4!, {r1,r6}
\r
1330 stmia r5!, {r1,r6}
\r
1331 bne .fl_loopcpRGB555_sh
\r
1333 @ hilighted pixels:
\r
1334 sub r3, r3, #0x40*2
\r
1336 .fl_loopcpRGB555_hi:
\r
1338 TileDoShHi2Pixels555 r1
\r
1341 bne .fl_loopcpRGB555_hi
\r
1343 sub r3, r3, #0x40*2
\r
1347 ldr r1, =(HighCol+8)
\r
1349 mov lr, lr, lsl #1
\r
1357 and r4, lr, r12, lsl #1
\r
1359 and r5, lr, r12, lsr #7
\r
1361 orr r4, r4, r5, lsl #16
\r
1363 and r5, lr, r12, lsr #15
\r
1365 and r6, lr, r12, lsr #23
\r
1367 orr r5, r5, r6, lsl #16
\r
1369 and r8, lr, r7, lsl #1
\r
1371 and r6, lr, r7, lsr #7
\r
1373 orr r8, r8, r6, lsl #16
\r
1375 and r12,lr, r7, lsr #15
\r
1376 ldrh r12,[r3, r12]
\r
1377 and r6, lr, r7, lsr #23
\r
1379 orr r12,r12, r6, lsl #16
\r
1381 stmia r0!, {r4,r5,r8,r12}
\r
1382 bne .fl_loopRGB555
\r
1385 ldmfd sp!, {r4-r8,lr}
\r
1388 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1391 .global blockcpy @ void *dst, void *src, size_t n
\r
1394 stmfd sp!, {r4,r5}
\r
1395 mov r2, r2, lsr #4
\r
1398 ldmia r1!, {r3-r5,r12}
\r
1399 stmia r0!, {r3-r5,r12}
\r
1401 ldmfd sp!, {r4,r5}
\r
1405 .global blockcpy_or @ void *dst, void *src, size_t n, int pat
\r
1408 stmfd sp!, {r4-r6}
\r
1409 orr r3, r3, r3, lsl #8
\r
1410 orr r3, r3, r3, lsl #16
\r
1411 mov r2, r2, lsr #4
\r
1414 ldmia r1!, {r4-r6,r12}
\r
1419 stmia r0!, {r4-r6,r12}
\r
1420 bne blockcpy_loop_or
\r
1421 ldmfd sp!, {r4-r6}
\r