2 * assembly optimized versions of most funtions from draw.c
\r
3 * (C) notaz, 2006-2010,2017
\r
5 * This work is licensed under the terms of MAME license.
\r
6 * See COPYING file in the top-level directory.
\r
8 * this is highly specialized, be careful if changing related C code!
\r
11 #include "pico_int_o32.h"
\r
13 .extern DrawStripInterlace
\r
15 .equ PDRAW_SPRITES_MOVED, (1<<0)
\r
16 .equ PDRAW_WND_DIFF_PRIO, (1<<1)
\r
17 .equ PDRAW_ACC_SPRITES, (1<<2)
\r
18 .equ PDRAW_DIRTY_SPRITES, (1<<4)
\r
19 .equ PDRAW_PLANE_HI_PRIO, (1<<6)
\r
20 .equ PDRAW_SHHI_DONE, (1<<7)
\r
23 .macro add_c24 d s c
\r
24 add \d, \s, #(\c & 0x00ff00)
\r
26 add \d, \d, #(\c & 0x0000ff)
\r
29 add \d, \d, #(\c & 0xff0000)
\r
33 .macro TilePixel pat lsrr offs
\r
37 ands r4, \pat, r2, lsr #\lsrr
\r
40 strneb r4, [r1,#\offs]
\r
43 @ TileNorm (r1=pdest, r2=pixels8, r3=pal) r4: scratch, pat: register with helper pattern 0xf
\r
45 TilePixel \pat, 12, 0 @ #0x0000f000
\r
46 TilePixel \pat, 8, 1 @ #0x00000f00
\r
47 TilePixel \pat, 4, 2 @ #0x000000f0
\r
48 TilePixel \pat, 0, 3 @ #0x0000000f
\r
49 TilePixel \pat, 28, 4 @ #0xf0000000
\r
50 TilePixel \pat, 24, 5 @ #0x0f000000
\r
51 TilePixel \pat, 20, 6 @ #0x00f00000
\r
52 TilePixel \pat, 16, 7 @ #0x000f0000
\r
55 @ TileFlip (r1=pdest, r2=pixels8, r3=pal) r4: scratch, pat: register with helper pattern 0xf
\r
57 TilePixel \pat, 16, 0 @ #0x000f0000
\r
58 TilePixel \pat, 20, 1 @ #0x00f00000
\r
59 TilePixel \pat, 24, 2 @ #0x0f000000
\r
60 TilePixel \pat, 28, 3 @ #0xf0000000
\r
61 TilePixel \pat, 0, 4 @ #0x0000000f
\r
62 TilePixel \pat, 4, 5 @ #0x000000f0
\r
63 TilePixel \pat, 8, 6 @ #0x00000f00
\r
64 TilePixel \pat, 12, 7 @ #0x0000f000
\r
67 @ shadow/hilight mode
\r
69 @ this one is for hi priority layer
\r
70 .macro TilePixelShHP lsrr offs
\r
74 ands r4, r12, r2, lsr #\lsrr
\r
76 ldreqb r4, [r1,#\offs]
\r
79 strb r4, [r1,#\offs]
\r
82 @ TileNormShHP (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: register with helper pattern 0xf, touches r3 high bits
\r
84 TilePixelShHP 12, 0 @ #0x0000f000
\r
85 TilePixelShHP 8, 1 @ #0x00000f00
\r
86 TilePixelShHP 4, 2 @ #0x000000f0
\r
87 TilePixelShHP 0, 3 @ #0x0000000f
\r
88 TilePixelShHP 28, 4 @ #0xf0000000
\r
89 TilePixelShHP 24, 5 @ #0x0f000000
\r
90 TilePixelShHP 20, 6 @ #0x00f00000
\r
91 TilePixelShHP 16, 7 @ #0x000f0000
\r
94 @ TileFlipShHP (r1=pdest, r2=pixels8, r3=pal) r4: scratch, pat: register with helper pattern 0xf
\r
96 TilePixelShHP 16, 0 @ #0x000f0000
\r
97 TilePixelShHP 20, 1 @ #0x00f00000
\r
98 TilePixelShHP 24, 2 @ #0x0f000000
\r
99 TilePixelShHP 28, 3 @ #0xf0000000
\r
100 TilePixelShHP 0, 4 @ #0x0000000f
\r
101 TilePixelShHP 4, 5 @ #0x000000f0
\r
102 TilePixelShHP 8, 6 @ #0x00000f00
\r
103 TilePixelShHP 12, 7 @ #0x0000f000
\r
107 @ TileSingleSh (r1=pdest, r2=pixels8, r3=pal) r4,r7: scratch, r0=sx; r12: helper pattern 0xf
\r
108 .macro TileSingleSh
\r
109 tst r0, #1 @ not aligned?
\r
115 strneb r4, [r1], #1
\r
116 streqh r4, [r1], #2
\r
128 strneb r4, [r1], #1
\r
131 @ TileSingleHi (r1=pdest, r2=pixels8, r3=pal) r4,r7: scratch, r0=sx, r12: register with helper pattern 0xf
\r
132 .macro TileSingleHi
\r
133 tst r1, #1 @ not aligned?
\r
136 ldrneb r4, [r1], #1
\r
137 ldreqh r4, [r1], #2 @ 1ci
\r
139 bic r4, r4, r7, lsr #1
\r
141 strneb r4, [r1, #-3]
\r
142 streqh r4, [r1, #-4]
\r
144 bic r12, r12, r7, lsr #1
\r
146 strh r12, [r1, #-4]
\r
148 bic r4, r4, r7, lsr #1
\r
152 bic r12, r12, r7, lsr #1
\r
154 strh r12, [r1, #-2]
\r
155 bicne r4, r4, r7, lsr #1
\r
157 strneb r4, [r1], #1
\r
161 .macro TileDoShGenPixel shift ofs
\r
163 ands r4, r12, r2, lsr #\shift
\r
169 ldrgeb r7, [r1,#\ofs]
\r
170 orrlt r7, r3, r4 @ normal
\r
172 bicge r7, r7, #0xc0
\r
173 orrge r7, r7, r4, lsl #6
\r
174 strb r7, [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_markop shift ofs
\r
204 ands r4, r12, r2, lsr #\shift
\r
210 ldrgeb r4, [r1,#\ofs]
\r
212 orrge r4, r4, #0x80
\r
213 strb r4, [r1,#\ofs]
\r
217 .macro TileFlipSh_markop
\r
218 TileDoShGenPixel_markop 16, 0 @ #0x000f0000
\r
219 TileDoShGenPixel_markop 20, 1 @ #0x00f00000
\r
220 TileDoShGenPixel_markop 24, 2 @ #0x0f000000
\r
221 TileDoShGenPixel_markop 28, 3 @ #0xf0000000
\r
222 TileDoShGenPixel_markop 0, 4 @ #0x0000000f
\r
223 TileDoShGenPixel_markop 4, 5 @ #0x000000f0
\r
224 TileDoShGenPixel_markop 8, 6 @ #0x00000f00
\r
225 TileDoShGenPixel_markop 12, 7 @ #0x0000f000
\r
228 .macro TileNormSh_markop
\r
229 TileDoShGenPixel_markop 12, 0 @ #0x0000f000
\r
230 TileDoShGenPixel_markop 8, 1 @ #0x00000f00
\r
231 TileDoShGenPixel_markop 4, 2 @ #0x000000f0
\r
232 TileDoShGenPixel_markop 0, 3 @ #0x0000000f
\r
233 TileDoShGenPixel_markop 28, 4 @ #0xf0000000
\r
234 TileDoShGenPixel_markop 24, 5 @ #0x0f000000
\r
235 TileDoShGenPixel_markop 20, 6 @ #0x00f00000
\r
236 TileDoShGenPixel_markop 16, 7 @ #0x000f0000
\r
239 .macro TileDoShGenPixel_onlyop_lp shift ofs
\r
241 ands r7, r12, r2, lsr #\shift
\r
245 ldrneb r4, [r1,#\ofs]
\r
250 bicne r4, r4, #0xc0
\r
251 orrne r4, r4, r7, lsl #6
\r
252 strneb r4, [r1,#\ofs]
\r
256 .macro TileFlipSh_onlyop_lp
\r
257 TileDoShGenPixel_onlyop_lp 16, 0 @ #0x000f0000
\r
258 TileDoShGenPixel_onlyop_lp 20, 1 @ #0x00f00000
\r
259 TileDoShGenPixel_onlyop_lp 24, 2 @ #0x0f000000
\r
260 TileDoShGenPixel_onlyop_lp 28, 3 @ #0xf0000000
\r
261 TileDoShGenPixel_onlyop_lp 0, 4 @ #0x0000000f
\r
262 TileDoShGenPixel_onlyop_lp 4, 5 @ #0x000000f0
\r
263 TileDoShGenPixel_onlyop_lp 8, 6 @ #0x00000f00
\r
264 TileDoShGenPixel_onlyop_lp 12, 7 @ #0x0000f000
\r
267 .macro TileNormSh_onlyop_lp
\r
268 TileDoShGenPixel_onlyop_lp 12, 0 @ #0x0000f000
\r
269 TileDoShGenPixel_onlyop_lp 8, 1 @ #0x00000f00
\r
270 TileDoShGenPixel_onlyop_lp 4, 2 @ #0x000000f0
\r
271 TileDoShGenPixel_onlyop_lp 0, 3 @ #0x0000000f
\r
272 TileDoShGenPixel_onlyop_lp 28, 4 @ #0xf0000000
\r
273 TileDoShGenPixel_onlyop_lp 24, 5 @ #0x0f000000
\r
274 TileDoShGenPixel_onlyop_lp 20, 6 @ #0x00f00000
\r
275 TileDoShGenPixel_onlyop_lp 16, 7 @ #0x000f0000
\r
279 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
283 @ int nametab; // 0x00
\r
284 @ int line; // 0x04
\r
285 @ int hscroll; // 0x08
\r
286 @ int xmask; // 0x0C
\r
287 @ int *hc; // 0x10 (pointer to cache buffer)
\r
288 @ int cells; // 0x14
\r
291 @ void DrawLayer(int lflags, int *hcache, int cellskip, int maxcells,
\r
292 @ struct PicoEState *est)
\r
297 ldr r12, [sp] @ est
\r
298 stmfd sp!, {r4-r11,lr}
\r
300 ldr r11, [r12, #OFS_EST_Pico]
\r
303 ldrb r7, [r11, #OFS_Pico_video_reg+16] @ ??vv??hh
\r
305 mov r6, r1 @ hcache
\r
306 orr r9, r3, r0, lsl #29 @ r9=force[31]|sh[30]|plane[29]
\r
307 orr r9, r9, r2, lsl #8 @ |cellskip[15:8]|maxcells[7:0] (tmp)
\r
310 orr r1, r1, #0x00ff
\r
314 biclt r1, r1, #0xfc00
\r
315 biceq r1, r1, #0xfe00
\r
318 movgt r1, #0x00ff @ r1=ymask=(height<<8)|0xff; ...; // Y Mask in pixels
\r
322 subge r10, r10, #1 @ r10=shift[width] (5,6,6,7)
\r
324 ldr r2, [r12, #OFS_EST_DrawScanline]
\r
325 ldr lr, [r12, #OFS_EST_PicoMem_vram]
\r
329 ldreqb r12, [r11, #OFS_Pico_video_reg+2]
\r
330 ldrneb r12, [r11, #OFS_Pico_video_reg+4]
\r
333 mov r5, r8, lsl r10
\r
334 sub r5, r5, #1 @ r5=xmask
\r
336 moveq r12, r12, lsl #10
\r
337 movne r12, r12, lsl #13
\r
338 and r12, r12, #(7<<13) @ r12=(ts->nametab<<1) (halfword compliant)
\r
340 ldrh r8, [r11, #OFS_Pico_video_reg+12]
\r
341 ldrb r7, [r11, #OFS_Pico_video_reg+11]
\r
343 mov r4, r8, lsr #8 @ pvid->reg[13]
\r
344 mov r4, r4, lsl #10 @ htab=pvid->reg[13]<<9; (halfwords)
\r
346 addne r4, r4, r2, lsl #2 @ htab+=DrawScanline<<1; // Offset by line
\r
348 biceq r4, r4, #0x1f @ htab&=~0xf; // Offset by tile
\r
349 add r4, r4, r0, lsl #1 @ htab+=plane
\r
350 bic r4, r4, #0x00ff0000 @ just in case
\r
351 ldrh r3, [lr, r4] @ r3=hscroll
\r
354 bne .DrawStrip_vsscroll
\r
356 @ Get vertical scroll value:
\r
357 add_c24 r7, lr, (OFS_PMEM_vsram-OFS_PMEM_vram)
\r
362 bne .DrawStrip_interlace
\r
365 movne r7, r7, lsr #16
\r
367 @ Find the line in the name table
\r
371 add r10, r10, #1 @ shift[width]++
\r
372 add r12, r12, r4, lsl r10 @ nametab+=(ts.line>>3)<<shift[width];
\r
374 @ ldmia r0, {r1,r2,r3,r5,r6,r9} @ r2=line, r3=ts->hscroll, r5=ts->xmask, r6=ts->hc, r9=ts->cells
\r
377 mov r10,r10, lsl #1 @ r10=ty=(ts->line&7)<<1;
\r
378 orr r10,r10, r9, lsl #24
\r
381 mov r8, r8, lsr #3 @ r8=tilex=(-ts->hscroll)>>3
\r
385 add r7, r1, #1 @ r7=dx=((ts->hscroll-1)&7)+1
\r
387 movs r3, r9, lsl #1 @ (force[31]|sh[30]) << 1
\r
389 orrmi r10,r10, #1<<23 @ r10=cells[31:24]|sh[23]|hi_not_empty[22]
\r
390 orrcs r10,r10, #1<<20 @ |had_output[21]|force[20]|ty[15:0]
\r
391 movmi r3, #0x40 @ default to shadowed pal on sh mode
\r
394 addne r10,r10, #0x01000000 @ we will loop cells+1 times if there is scroll
\r
396 and r9, r9, #0xff00
\r
397 add r8, r8, r9, lsr #8 @ tilex+=cellskip
\r
398 add r7, r7, r9, lsr #5 @ dx+=cellskip<<3;
\r
399 sub r10,r10,r9, lsl #16 @ cells-=cellskip
\r
401 @ cache some stuff to avoid mem access
\r
402 ldr r11,[sp, #9*4] @ est
\r
404 ldr r11,[r11, #OFS_EST_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 tsteq r10, #1<<20 @ force?
\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, #0x00d00000 @ sh[23]|hi_not_empty[22]|force[20]
\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 ldr r1, [sp, #9*4] @ est
\r
509 ldreq r2, [r1, #OFS_EST_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
512 streq r2, [r1, #OFS_EST_rendstatus]
\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 r11, [sp, #9*4] @ est
\r
526 orr r5, r1, r10, lsl #24
\r
527 ldr r4, [r11, #OFS_EST_DrawScanline]
\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
534 orr r10,r10, #0xff000000 @ will be adjusted on entering loop
\r
536 orrne r10,r10, #0x8000
\r
537 tst r3, #0x0f @ hscroll & 0x0f?
\r
540 sub r10,r10, #1<<24 @ cell-- // start from negative for hscroll
\r
542 subne r10,r10, #1<<16 @ cells--
\r
543 subne r10,r10, #1<<24 @ cell-- // even more negative
\r
547 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
548 movne r3, #0x40 @ default to shadowed pal on sh mode
\r
550 and r9, r9, #0xff00
\r
551 add r8, r8, r9, lsr #8 @ tilex+=cellskip
\r
552 add r7, r7, r9, lsr #5 @ dx+=cellskip<<3;
\r
553 add r10,r10,r9, lsl #16 @ cell+=cellskip
\r
555 @ cache some stuff to avoid mem access
\r
556 ldr r11,[sp, #9*4] @ est
\r
558 ldr r11,[r11, #OFS_EST_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_c24 r7, lr, (OFS_PMEM_vsram-OFS_PMEM_vram)
\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 @ PicoMem.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 *)(PicoMem.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 *)(PicoMem.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 ldr r1, [sp, #9*4] @ est
\r
683 ldreq r2, [r1, #OFS_EST_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
686 streq r2, [r1, #OFS_EST_rendstatus]
\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
724 @ void BackFill(int reg7, int sh, struct PicoEState *est)
\r
729 stmfd sp!, {r4-r9,lr}
\r
731 mov r0, r0, lsl #26
\r
732 ldr lr, [r2, #OFS_EST_HighCol]
\r
733 mov r0, r0, lsr #26
\r
736 orr r0, r0, r1, lsl #6
\r
737 orr r0, r0, r0, lsl #8
\r
738 orr r0, r0, r0, lsl #16
\r
749 stmia lr!, {r0-r7} @ 10*8*4
\r
760 ldmfd sp!, {r4-r9,lr}
\r
764 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
766 @ void DrawTilesFromCache(int *hc, int sh, int rlim, struct PicoEState *est)
\r
768 .global DrawTilesFromCache
\r
770 DrawTilesFromCache:
\r
771 stmfd sp!, {r4-r9,r11,lr}
\r
773 @ cache some stuff to avoid mem access
\r
774 ldr r11,[r3, #OFS_EST_HighCol]
\r
776 ldr lr, [r3, #OFS_EST_PicoMem_vram]
\r
779 mvn r5, #0 @ r5=prevcode=-1
\r
781 orr r8, r8, r2, lsl #1
\r
782 bne .dtfc_check_rendflags
\r
786 ldr r6, [r0], #4 @ read code
\r
787 movs r1, r6, lsr #16 @ r1=dx;
\r
788 ldmeqfd sp!, {r4-r9,r11,pc} @ dx is never zero, this must be a terminator, return
\r
789 bic r4, r1, #0xfe00
\r
790 add r1, r11, r4 @ r1=pdest
\r
792 movs r7, r6, lsl #16
\r
793 bpl .dtfc_loop @ !(code & 0x8000)
\r
794 cmp r5, r7, lsr #16
\r
795 beq .dtfc_samecode @ if (code==prevcode)
\r
797 mov r5, r7, lsr #16
\r
799 mov r2, r5, lsl #21
\r
800 mov r2, r2, lsr #17 @ r2=addr=(code&0x7ff)<<4;
\r
801 add r2, r2, r6, lsr #25 @ addr+=ty
\r
803 and r3, r5, #0x6000
\r
804 mov r3, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);
\r
806 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
809 rsbs r4, r4, r8, lsr #1
\r
819 beq .dtfc_SingleColor @ tileline singlecolor
\r
824 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
836 orr r4, r4, r4, lsl #8
\r
837 tst r1, #1 @ not aligned?
\r
838 strneb r4, [r1], #1
\r
839 streqh r4, [r1], #2
\r
843 strneb r4, [r1], #1 @ have a remaining unaligned pixel?
\r
848 beq .dtfc_shadow_blank
\r
851 beq .dtfc_SingleColor @ tileline singlecolor
\r
854 bne .dtfc_TileFlipShHP
\r
856 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
857 .dtfc_TileNormShHP:
\r
861 .dtfc_TileFlipShHP:
\r
865 .dtfc_shadow_blank:
\r
870 strneb r4, [r1], #1
\r
872 orr r6, r6, r6, lsl #8
\r
888 add r4, r4, #7 @ 0-6
\r
892 mov r2, r2, ror #16
\r
893 tst r5, #0x0800 @ flipped?
\r
896 mov r2, r2, ror #16
\r
906 @ check if we have detected layer covered with hi-prio tiles:
\r
907 .dtfc_check_rendflags:
\r
908 ldr r2, [r9, #OFS_EST_rendstatus]
\r
909 tst r2, #(PDRAW_PLANE_HI_PRIO|PDRAW_SHHI_DONE)
\r
911 bic r8, r8, #1 @ sh/hi mode off
\r
912 tst r2, #PDRAW_SHHI_DONE
\r
913 bne .dtfc_loop @ already processed
\r
914 orr r2, r2, #PDRAW_SHHI_DONE
\r
915 str r2, [r9, #OFS_EST_rendstatus]
\r
920 orr r6, r6, r6, lsl #8
\r
921 orr r6, r6, r6, lsl #16
\r
923 ldmia r1, {r2,r4,r5,r7}
\r
929 stmia r1!,{r2,r4,r5,r7}
\r
930 bne .dtfc_loop_shprep
\r
932 mvn r5, #0 @ r5=prevcode=-1
\r
937 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
940 @ void DrawSpritesSHi(unsigned char *sprited, struct PicoEState *est)
\r
942 .global DrawSpritesSHi
\r
950 stmfd sp!, {r1,r4-r11,lr} @ +est
\r
951 strb r12,[r0,#2] @ set end marker
\r
952 add r10,r0, #3 @ r10=HighLnSpr end
\r
953 add r10,r10,r3 @ r10=HighLnSpr end
\r
955 ldr r11,[r1, #OFS_EST_HighCol]
\r
957 ldr lr, [r1, #OFS_EST_PicoMem_vram]
\r
962 ldrb r0, [r10,#-1]!
\r
964 ldr r1, [r7, #OFS_EST_HighPreSpr]
\r
966 ldmeqfd sp!, {r1,r4-r11,pc} @ end of list
\r
968 add r0, r1, r0, lsl #3
\r
970 ldr r9, [r0, #4] @ sprite[1]
\r
971 mov r2, r9, asr #16 @ r2=sx
\r
973 mov r9, r9, lsl #16
\r
974 mov r3, r9, lsr #31 @ priority
\r
975 mov r9, r9, lsr #16
\r
976 @ orr r9, r9, r8, lsl #31 @ r9=code|sh[31] @@ sh is always on here now
\r
977 and r4, r9, #0x6000
\r
978 orr r9, r9, r4, lsl #16
\r
979 orr r9, r9, #0x90000000 @ r9=scc1 ???? ... <code> (s=shadow/hilight, cc=pal)
\r
980 cmp r12,r9, lsr #28 @ sh/hi with pal3?
\r
981 cmpne r3, #1 @ if not, is it hi prio?
\r
982 bne DrawSpriteSHi @ non-operator low sprite, already drawn
\r
984 ldr r3, [r0] @ sprite[0]
\r
985 mov r6, r3, lsr #28
\r
986 sub r6, r6, #1 @ r6=width-1 (inc later)
\r
987 mov r5, r3, lsr #24
\r
988 and r5, r5, #7 @ r5=height
\r
990 ldr r7, [r7, #OFS_EST_DrawScanline]
\r
991 mov r0, r3, lsl #16 @ r4=sy<<16 (tmp)
\r
993 sub r7, r7, r0, asr #16 @ r7=row=DrawScanline-sy
\r
996 movne r0, r5, lsl #3
\r
998 subne r7, r0, r7 @ if (code&0x1000) row=(height<<3)-1-row; // Flip Y
\r
1000 add r8, r9, r7, lsr #3 @ tile+=row>>3; // Tile number increases going down
\r
1002 mlane r8, r5, r6, r8 @ if (code&0x0800) { tile+=delta*(width-1);
\r
1003 rsbne r5, r5, #0 @ delta=-delta; } // r5=delta now
\r
1005 mov r8, r8, lsl #21
\r
1006 mov r8, r8, lsr #17
\r
1008 add r8, r8, r7, lsl #1 @ tile+=(row&7)<<1; // Tile address
\r
1010 mov r5, r5, lsl #4 @ delta<<=4; // Delta of address
\r
1011 mov r3, r4, lsr #9 @ r3=pal=((code>>9)&0x30);
\r
1013 add r6, r6, #1 @ inc now
\r
1014 adds r0, r2, #0 @ mov sx to r0 and set ZV flags
\r
1015 b .dsprShi_loop_enter
\r
1018 subs r6, r6, #1 @ width--
\r
1020 adds r0, r0, #8 @ sx+=8
\r
1021 add r8, r8, r5 @ tile+=delta
\r
1023 .dsprShi_loop_enter:
\r
1024 ble .dsprShi_loop @ sx <= 0
\r
1028 mov r8, r8, lsl #17
\r
1029 mov r8, r8, lsr #17 @ tile&=0x7fff; // Clip tile address
\r
1031 ldr r2, [lr, r8, lsl #1] @ pack=*(unsigned int *)(PicoMem.vram+addr); // Get 8 pixels
\r
1032 add r1, r11, r0 @ r1=pdest
\r
1036 cmp r12, r9, lsr #28
\r
1037 beq .dsprShi_shadow
\r
1039 cmp r2, r2, ror #4
\r
1040 beq .dsprShi_SingleColor @ tileline singlecolor
\r
1043 bne .dsprShi_TileFlip
\r
1045 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
1047 .dsprShi_TileNorm:
\r
1051 .dsprShi_TileFlip:
\r
1055 .dsprShi_SingleColor:
\r
1058 orr r4, r4, r4, lsl #8
\r
1059 tst r0, #1 @ not aligned?
\r
1060 strneb r4, [r1], #1
\r
1061 streqh r4, [r1], #2
\r
1065 strneb r4, [r1], #1
\r
1070 beq .dsprShi_shadow_lowpri
\r
1072 cmp r2, r2, ror #4
\r
1073 beq .dsprShi_singlec_sh
\r
1076 bne .dsprShi_TileFlip_sh
\r
1078 @ (r1=pdest, r2=pixels8, r3=pal) r4, r7: scratch, r12: helper pattern
\r
1079 .dsprShi_TileNorm_sh:
\r
1083 .dsprShi_TileFlip_sh:
\r
1087 .dsprShi_singlec_sh:
\r
1088 cmp r2, #0xe0000000
\r
1089 bcc .dsprShi_SingleColor @ normal singlecolor tileline (carry inverted in ARM)
\r
1090 tst r2, #0x10000000
\r
1091 bne .dsprShi_sh_sh
\r
1099 .dsprShi_shadow_lowpri:
\r
1101 bne .dsprShi_TileFlip_sh_lp
\r
1103 .dsprShi_TileNorm_sh_lp:
\r
1104 TileNormSh_onlyop_lp
\r
1107 .dsprShi_TileFlip_sh_lp:
\r
1108 TileFlipSh_onlyop_lp
\r
1113 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1115 @ void DrawAllSprites(unsigned char *sprited, int prio, int sh,
\r
1116 @ struct PicoEState *est)
\r
1118 .global DrawAllSprites
\r
1121 orr r1, r2, r1, lsl #1
\r
1123 ands r2, r2, #0x7f
\r
1126 @ time to do some real work
\r
1127 stmfd sp!, {r1,r3-r11,lr} @ +sh|prio<<1 +est
\r
1129 strb r12,[r0,#2] @ set end marker
\r
1131 add r10,r10,r2 @ r10=HighLnSpr end
\r
1133 ldr r11,[r3, #OFS_EST_HighCol]
\r
1135 ldr lr, [r3, #OFS_EST_PicoMem_vram]
\r
1137 @ + 0 : hhhhvvvv ----hhvv yyyyyyyy yyyyyyyy // v, h: horiz. size
\r
1138 @ + 4 : xxxxxxxx xxxxxxxx pccvhnnn nnnnnnnn // x: x coord + 8
\r
1141 @ draw next sprite
\r
1142 ldrb r0, [r10,#-1]!
\r
1143 ldr r8, [sp] @ sh|prio<<1
\r
1144 ldr r7, [sp, #4] @ est
\r
1145 mov r2, r0, lsr #7
\r
1147 ldmeqfd sp!, {r1,r3-r11,pc} @ end of list
\r
1148 cmp r2, r8, lsr #1
\r
1149 bne DrawSprite @ wrong priority
\r
1150 ldr r1, [r7, #OFS_EST_HighPreSpr]
\r
1152 add r0, r1, r0, lsl #3
\r
1154 ldr r3, [r0] @ sprite[0]
\r
1155 ldr r7, [r7, #OFS_EST_DrawScanline]
\r
1156 mov r6, r3, lsr #28
\r
1157 sub r6, r6, #1 @ r6=width-1 (inc later)
\r
1158 mov r5, r3, lsr #24
\r
1159 and r5, r5, #7 @ r5=height
\r
1161 mov r4, r3, lsl #16 @ r4=sy<<16 (tmp)
\r
1164 sub r7, r7, r4, asr #16 @ r7=row=DrawScanline-sy
\r
1166 mov r2, r9, asr #16 @ r2=sx
\r
1167 mov r9, r9, lsl #16
\r
1168 mov r9, r9, lsr #16
\r
1169 orr r9, r9, r8, lsl #31 @ r9=code|sh[31]
\r
1172 movne r4, r5, lsl #3
\r
1174 subne r7, r4, r7 @ if (code&0x1000) row=(height<<3)-1-row; // Flip Y
\r
1176 add r8, r9, r7, lsr #3 @ tile+=row>>3; // Tile number increases going down
\r
1178 mlane r8, r5, r6, r8 @ if (code&0x0800) { tile+=delta*(width-1);
\r
1179 rsbne r5, r5, #0 @ delta=-delta; } // r5=delta now
\r
1181 mov r8, r8, lsl #21
\r
1182 mov r8, r8, lsr #17
\r
1184 add r8, r8, r7, lsl #1 @ tile+=(row&7)<<1; // Tile address
\r
1187 @ cache some stuff to avoid mem access
\r
1188 mov r5, r5, lsl #4 @ delta<<=4; // Delta of address
\r
1189 and r4, r9, #0x6000
\r
1190 orr r9, r9, r4, lsl #16
\r
1191 orrs r9, r9, #0x10000000 @ r9=scc1 ???? ... <code> (s=shadow/hilight, cc=pal)
\r
1193 mov r3, r4, lsr #9 @ r3=pal=((code>>9)&0x30);
\r
1194 orrmi r3, r3, #0x40 @ for sh/hi
\r
1196 add r6, r6, #1 @ inc now
\r
1197 adds r0, r2, #0 @ mov sx to r0 and set ZV flags
\r
1198 b .dspr_loop_enter
\r
1201 subs r6, r6, #1 @ width--
\r
1203 adds r0, r0, #8 @ sx+=8
\r
1204 add r8, r8, r5 @ tile+=delta
\r
1207 ble .dspr_loop @ sx <= 0
\r
1211 mov r8, r8, lsl #17
\r
1212 mov r8, r8, lsr #17 @ tile&=0x7fff; // Clip tile address
\r
1214 ldr r2, [lr, r8, lsl #1] @ pack=*(unsigned int *)(PicoMem.vram+addr); // Get 8 pixels
\r
1215 add r1, r11, r0 @ r1=pdest
\r
1219 cmp r12, r9, lsr #28
\r
1222 cmp r2, r2, ror #4
\r
1223 beq .dspr_SingleColor @ tileline singlecolor
\r
1226 bne .dspr_TileFlip
\r
1228 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
1239 cmp r2, #0xe0000000
\r
1240 bcs .dspr_TileNorm_sh @ op. tileline, markop. XXX: maybe add a spec. handler?
\r
1242 .dspr_SingleColor:
\r
1245 orr r4, r4, r4, lsl #8
\r
1246 tst r0, #1 @ not aligned?
\r
1247 strneb r4, [r1], #1
\r
1248 streqh r4, [r1], #2
\r
1252 strneb r4, [r1], #1
\r
1256 cmp r2, r2, ror #4
\r
1257 beq .dspr_singlec_sh
\r
1260 bne .dspr_TileFlip_sh
\r
1262 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern
\r
1263 .dspr_TileNorm_sh:
\r
1267 .dspr_TileFlip_sh:
\r
1272 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1274 @ void DrawWindow(int tstart, int tend, int prio, int sh
\r
1275 @ struct PicoEState *est)
\r
1277 .global DrawWindow
\r
1280 ldr r12, [sp] @ est
\r
1281 stmfd sp!, {r4-r11,lr}
\r
1283 ldr r6, [r12, #OFS_EST_Pico]
\r
1284 ldr r10, [r12, #OFS_EST_DrawScanline]
\r
1285 mov r11, r12 @ est
\r
1286 ldrb r12, [r6, #OFS_Pico_video_reg+3] @ pvid->reg[3]
\r
1288 ldr r4, [r6, #OFS_Pico_video_reg+12]
\r
1289 mov r5, r10, lsr #3
\r
1291 mov r10, r10, lsl #1 @ r10=ty
\r
1293 mov r12, r12, lsl #10
\r
1295 tst r4, #1 @ 40 cell mode?
\r
1296 andne r12, r12, #0xf000 @ 0x3c<<10
\r
1297 andeq r12, r12, #0xf800
\r
1298 addne r12, r12, r5, lsl #7
\r
1299 addeq r12, r12, r5, lsl #6 @ nametab
\r
1300 add r12, r12, r0, lsl #2 @ +starttile
\r
1302 ldr lr, [r11, #OFS_EST_PicoMem_vram]
\r
1303 ldr r6, [r11, #OFS_EST_rendstatus]
\r
1305 @ fetch the first code now
\r
1306 ldrh r7, [lr, r12]
\r
1308 ands r6, r6, #PDRAW_WND_DIFF_PRIO
\r
1311 eoreq r8, r2, r7, lsr #15 @ do prio bits differ?
\r
1313 ldmeqfd sp!, {r4-r11,pc} @ yes, assume that whole window uses same priority
\r
1315 orr r6, r6, r3, lsl #8 @ shadow mode
\r
1319 @ cache some stuff to avoid mem access
\r
1320 ldr r11, [r11, #OFS_EST_HighCol]
\r
1321 mov r8, r8, lsl #1 @ cells
\r
1323 mvn r9, #0 @ r9=prevcode=-1
\r
1324 add r1, r11, r0, lsl #4 @ r1=pdest
\r
1328 @ r4,r5 are scratch in this loop
\r
1332 add r12, r12, #2 @ halfwords
\r
1333 ldrh r7, [lr, r12] @ r7=code (int, but from unsigned, no sign extend)
\r
1335 beq .dwloop_end @ done
\r
1337 eor r5, r6, r7, lsr #15
\r
1339 orrne r6, r6, #2 @ wrong pri
\r
1343 beq .dw_samecode @ we know stuff about this tile already
\r
1346 mov r9, r7 @ remember code
\r
1348 movs r2, r9, lsl #20 @ if (code&0x1000)
\r
1349 mov r2, r2, lsl #1
\r
1350 add r2, r10, r2, lsr #17 @ r2=addr=(code&0x7ff)<<4; addr+=ty
\r
1351 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe;
\r
1353 and r3, r9, #0x6000
\r
1354 mov r3, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);
\r
1356 ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels
\r
1363 beq .dwloop @ tileline blank
\r
1365 cmp r2, r2, ror #4
\r
1366 beq .dw_SingleColor @ tileline singlecolor
\r
1371 @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r0: helper pattern
\r
1381 and r4, r0, r2 @ #0x0000000f
\r
1383 orr r4, r4, r4, lsl #8
\r
1384 orr r4, r4, r4, lsl #16
\r
1386 stmia r1!, {r4,r5}
\r
1387 b .dwloop_nor1 @ we incremeted r1 ourselves
\r
1390 tst r6, #1 @ hi pri?
\r
1391 orreq r3, r3, #0x40
\r
1392 beq .dw_shadow_done
\r
1395 orr r5, r5, r5, lsl #8
\r
1396 orr r5, r5, r5, lsl #16
\r
1405 and r2, r6, #PDRAW_WND_DIFF_PRIO
\r
1406 ldmfd sp!, {r4-r11,lr}
\r
1408 ldr r1, [r0, #OFS_EST_rendstatus]
\r
1410 str r1, [r0, #OFS_EST_rendstatus]
\r
1415 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1418 @ hilights 2 pixels in RGB444/BGR444 format
\r
1419 .macro TileDoShHi2Pixels444 reg
\r
1420 mov \reg, \reg, ror #12
\r
1421 adds \reg, \reg, #0x40000000
\r
1422 orrcs \reg, \reg, #0xf0000000
\r
1423 mov \reg, \reg, ror #28
\r
1424 adds \reg, \reg, #0x40000000
\r
1425 orrcs \reg, \reg, #0xf0000000
\r
1426 mov \reg, \reg, ror #28
\r
1427 adds \reg, \reg, #0x40000000
\r
1428 orrcs \reg, \reg, #0xf0000000
\r
1429 mov \reg, \reg, ror #24
\r
1430 adds \reg, \reg, #0x40000000
\r
1431 orrcs \reg, \reg, #0xf0000000
\r
1432 mov \reg, \reg, ror #28
\r
1433 adds \reg, \reg, #0x40000000
\r
1434 orrcs \reg, \reg, #0xf0000000
\r
1435 mov \reg, \reg, ror #28
\r
1436 adds \reg, \reg, #0x40000000
\r
1437 orrcs \reg, \reg, #0xf0000000
\r
1438 mov \reg, \reg, ror #12
\r
1442 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1445 @ Convert 0000bbb0 ggg0rrr0
\r
1446 @ to rrrrrggg gggbbbbb
\r
1448 @ r2,r3 - scratch, lr = 0x001c001c, r8 = 0x08610861
\r
1449 .macro convRGB565 reg
\r
1450 and r2, lr, \reg,lsr #7 @ b
\r
1451 and r3, lr, \reg,lsr #3 @ g
\r
1452 and \reg, lr, \reg,lsl #1 @ r
\r
1453 orr r2, r2, r3, lsl #6
\r
1454 orr \reg, r2, \reg,lsl #11
\r
1456 and r2, r8, \reg,lsr #4
\r
1457 orr \reg, \reg, r2
\r
1460 @ trashes: r2-r8,r12,lr; r8 = 0x08610861; r0,r1 are advanced
\r
1461 .macro vidConvCpyRGB565_local
\r
1462 mov r12, r2, lsr #3 @ repeats
\r
1463 mov lr, #0x001c0000
\r
1464 orr lr, lr, #0x01c @ lr == pattern 0x001c001c
\r
1467 ldmia r1!, {r4-r7}
\r
1482 .global vidConvCpyRGB565
\r
1484 vidConvCpyRGB565: @ void *to, void *from, int pixels
\r
1485 stmfd sp!, {r4-r9,lr}
\r
1487 orr r8, r8, #0x0800
\r
1488 orr r8, r8, r8, lsl #16
\r
1489 vidConvCpyRGB565_local
\r
1490 ldmfd sp!, {r4-r9,lr}
\r
1494 @ void PicoDoHighPal555(int sh, int line, struct PicoEState *est)
\r
1496 .global PicoDoHighPal555
\r
1499 stmfd sp!, {r4-r10,lr}
\r
1502 ldr r8, [r10, #OFS_EST_Pico]
\r
1504 PicoDoHighPal555_nopush:
\r
1505 orr r9, r1, r0, lsl #31 @ 0:called from FinalizeLine555, 31: s/h
\r
1507 add r0, r10, #OFS_EST_HighPal
\r
1510 strb r1, [r8, #OFS_Pico_m_dirtyPal]
\r
1512 ldr r1, [r10, #OFS_EST_PicoMem_cram]
\r
1515 orr r8, r8, #0x0800
\r
1516 orr r8, r8, r8, lsl #16
\r
1518 vidConvCpyRGB565_local
\r
1521 beq PicoDoHighPal555_end
\r
1523 add r3, r10, #OFS_EST_HighPal
\r
1525 @ shadowed pixels:
\r
1527 add r4, r3, #0x40*2
\r
1528 orr r12,r12,#0x7300
\r
1529 add r5, r3, #0xc0*2
\r
1530 orr r12,r12,r12,lsl #16
\r
1532 .fl_loopcpRGB555_sh:
\r
1533 ldmia r3!, {r1,r6}
\r
1535 and r1, r12, r1, lsr #1
\r
1536 and r6, r12, r6, lsr #1
\r
1537 stmia r4!, {r1,r6}
\r
1538 stmia r5!, {r1,r6}
\r
1539 bne .fl_loopcpRGB555_sh
\r
1541 @ hilighted pixels:
\r
1542 @ t = ((dpal[i] >> 1) & 0x738e738e) + 0x738e738e;
\r
1543 @ t |= (t >> 4) & 0x08610861;
\r
1545 sub r3, r3, #0x40*2
\r
1547 .fl_loopcpRGB555_hi:
\r
1548 ldmia r3!, {r1,r6}
\r
1549 and r1, r12, r1, lsr #1
\r
1550 and r6, r12, r6, lsr #1
\r
1553 and r5, r8, r1, lsr #4
\r
1554 and r7, r8, r6, lsr #4
\r
1557 stmia r4!, {r1,r6}
\r
1559 bne .fl_loopcpRGB555_hi
\r
1562 PicoDoHighPal555_end:
\r
1564 ldmeqfd sp!, {r4-r10,pc}
\r
1566 ldr r8, [r10, #OFS_EST_Pico]
\r
1567 b FinalizeLineRGB555_pal_done
\r
1570 @ void FinalizeLine555(int sh, int line, struct PicoEState *est)
\r
1572 .global FinalizeLine555
\r
1575 stmfd sp!, {r4-r10,lr}
\r
1577 ldr r8, [r10, #OFS_EST_Pico]
\r
1579 ldrb r2, [r8, #OFS_Pico_m_dirtyPal]
\r
1582 bne PicoDoHighPal555_nopush
\r
1584 FinalizeLineRGB555_pal_done:
\r
1585 add r3, r10, #OFS_EST_HighPal
\r
1587 ldr r12, [r10, #OFS_EST_rendstatus]
\r
1588 eors r0, r0, #1 @ sh is 0
\r
1590 tstne r12,#PDRAW_ACC_SPRITES
\r
1593 ldr r1, [r10, #OFS_EST_HighCol]
\r
1594 ldr r0, [r10, #OFS_EST_DrawLineDest]
\r
1597 ldrb r12, [r8, #OFS_Pico_video_reg+12]
\r
1598 mov lr, lr, lsl #1
\r
1601 movne r2, #320/8 @ len
\r
1602 bne .fl_no32colRGB555
\r
1603 ldr r4, [r10, #OFS_EST_PicoOpt]
\r
1607 bne .fl_32scale_RGB555
\r
1609 addeq r0, r0, #32*2
\r
1611 .fl_no32colRGB555:
\r
1613 #ifdef UNALIGNED_DRAWLINEDEST
\r
1614 @ this is basically for Gizmondo, which has unaligned odd lines in the framebuffer
\r
1623 and r4, lr, r12, lsl #1
\r
1625 and r5, lr, r12, lsr #7
\r
1627 and r6, lr, r12, lsr #15
\r
1629 orr r4, r4, r5, lsl #16
\r
1631 and r5, lr, r12, lsr #23
\r
1633 and r8, lr, r7, lsl #1
\r
1635 orr r5, r6, r5, lsl #16
\r
1637 and r6, lr, r7, lsr #7
\r
1639 and r12,lr, r7, lsr #15
\r
1640 ldrh r12,[r3, r12]
\r
1641 and r7, lr, r7, lsr #23
\r
1643 orr r8, r8, r6, lsl #16
\r
1646 orr r12,r12, r7, lsl #16
\r
1648 stmia r0!, {r4,r5,r8,r12}
\r
1649 bne .fl_loopRGB555
\r
1651 ldmfd sp!, {r4-r10,lr}
\r
1655 .fl_32scale_RGB555:
\r
1656 mov r9, #0x3900 @ f800 07e0 001f | e000 0780 001c | 3800 01e0 0007
\r
1657 orr r9, r9, #0x00e7
\r
1659 #ifdef UNALIGNED_DRAWLINEDEST
\r
1661 bne .fl_32scale_RGB555u
\r
1664 .fl_loop32scale_RGB555:
\r
1668 and r4, lr, r12,lsl #1
\r
1670 and r5, lr, r12,lsr #7
\r
1672 and r4, r4, r9, lsl #2
\r
1673 orr r4, r4, r4, lsl #14 @ r4[31:16] = 1/4 pix_s 0
\r
1674 and r5, r5, r9, lsl #2
\r
1675 sub r6, r5, r5, lsr #2 @ r6 = 3/4 pix_s 1
\r
1676 add r4, r4, r6, lsl #16 @ pix_d 0, 1
\r
1677 and r6, lr, r12,lsr #15
\r
1679 and r12,lr, r12,lsr #23
\r
1680 ldrh r12,[r3, r12]
\r
1681 and r6, r6, r9, lsl #2
\r
1683 mov r5, r5, lsr #1
\r
1684 sub r6, r6, r6, lsr #2 @ r6 = 3/4 pix_s 2
\r
1685 orr r5, r5, r6, lsl #16
\r
1687 and r6, lr, r7, lsl #1
\r
1689 and r12,r12,r9, lsl #2
\r
1690 add r5, r5, r12,lsl #14 @ pix_d 2, 3
\r
1691 and r6, r6, r9, lsl #2
\r
1692 orr r6, r12,r6, lsl #16 @ pix_d 4, 5
\r
1694 and r12,lr, r7, lsr #7
\r
1695 ldrh r12,[r3, r12]
\r
1696 and r10,lr, r7, lsr #15
\r
1697 ldrh r10,[r3, r10]
\r
1698 and r12,r12,r9, lsl #2
\r
1699 sub r8, r12,r12,lsr #2 @ r8 = 3/4 pix_s 1
\r
1700 add r8, r8, r6, lsr #18
\r
1701 and r7, lr, r7, lsr #23
\r
1703 and r10,r10,r9, lsl #2
\r
1704 orr r8, r8, r10,lsl #15
\r
1705 add r8, r8, r12,lsl #15 @ pix_d 6, 7
\r
1706 sub r10,r10,r10,lsr #2 @ r10= 3/4 pix_s 2
\r
1707 and r7, r7, r9, lsl #2
\r
1708 add r10,r10,r7, lsr #2 @ += 1/4 pix_s 3
\r
1709 orr r10,r10,r7, lsl #16 @ pix_d 8, 9
\r
1713 stmia r0!, {r4,r5,r6,r8,r10}
\r
1714 bne .fl_loop32scale_RGB555
\r
1716 ldmfd sp!, {r4-r10,lr}
\r
1719 #ifdef UNALIGNED_DRAWLINEDEST
\r
1720 @ unaligned versions of loops
\r
1721 @ warning: starts drawing 2bytes before dst
\r
1724 sub r0, r0, #2 @ initial adjustment
\r
1731 and r6, lr, r12,lsl #1
\r
1733 and r5, lr, r12,lsr #7
\r
1735 orr r4, r8, r6, lsl #16
\r
1737 and r6, lr, r12,lsr #15
\r
1739 and r8, lr, r12,lsr #23
\r
1741 orr r5, r5, r6, lsl #16
\r
1743 and r6, lr, r7, lsl #1
\r
1745 and r12,lr, r7, lsr #7
\r
1746 ldrh r12,[r3, r12]
\r
1747 orr r6, r8, r6, lsl #16
\r
1749 and r8, lr, r7, lsr #15
\r
1751 and r7, lr, r7, lsr #23
\r
1754 orr r12,r12,r8, lsl #16
\r
1757 stmia r0!, {r4,r5,r6,r12}
\r
1758 bne .fl_loopRGB555u
\r
1762 ldmfd sp!, {r4-r10,lr}
\r
1766 .fl_32scale_RGB555u:
\r
1767 sub r0, r0, #2 @ initial adjustment
\r
1770 @ r9 f800 07e0 001f | e000 0780 001c | 3800 01e0 0007
\r
1771 .fl_loop32scale_RGB555u:
\r
1775 and r6, lr, r12,lsl #1
\r
1777 and r5, lr, r12,lsr #7
\r
1779 and r6, r6, r9, lsl #2
\r
1780 orr r4, r4, r6, lsl #16 @ r4 = pix_d -1, 0
\r
1782 and r5, r5, r9, lsl #2
\r
1783 sub r8, r5, r5, lsr #2 @ r8 = 3/4 pix_s 1
\r
1784 add r6, r8, r6, lsr #2 @ r6 = (1/4 pix_s 0) + (3/4 pix_s 1)
\r
1785 orr r5, r6, r5, lsl #15
\r
1787 and r6, lr, r12,lsr #15
\r
1789 and r12,lr, r12,lsr #23
\r
1790 ldrh r12,[r3, r12]
\r
1791 and r6, r6, r9, lsl #2
\r
1792 add r5, r5, r6, lsl #15 @ r5 = pix_d 1, 2
\r
1794 and r8, lr, r7, lsl #1
\r
1796 and r10,lr, r7, lsr #7
\r
1797 ldrh r10,[r3, r10]
\r
1798 and r12,r12,r9, lsl #2
\r
1799 sub r6, r6, r6, lsr #2 @ r6 = 3/4 pix_s 2
\r
1800 add r6, r6, r12,lsr #2
\r
1801 orr r6, r6, r12,lsl #16 @ r6 = pix_d 3, 4
\r
1803 and r8, r8, r9, lsl #2
\r
1804 and r10,r10,r9, lsl #2
\r
1805 sub r12,r10,r10,lsr #2 @ r12 = 3/4 pix_s 5
\r
1806 orr r8, r8, r8, lsl #14
\r
1807 add r8, r8, r12,lsl #16 @ r8 = pix_d 5, 6
\r
1808 and r12,lr, r7, lsr #15
\r
1809 ldrh r12,[r3, r12]
\r
1810 and r7, lr, r7, lsr #23
\r
1812 and r12,r12,r9, lsl #2
\r
1814 mov r10,r10, lsr #1
\r
1815 sub r12,r12,r12,lsr #2 @ r12 = 3/4 pix_s 6
\r
1816 orr r10,r10,r12,lsl #16
\r
1817 and r7, r7, r9, lsl #2
\r
1818 add r10,r10,r7, lsl #14 @ r10 = pix_d 7, 8
\r
1822 stmia r0!, {r4,r5,r6,r8,r10}
\r
1824 bne .fl_loop32scale_RGB555u
\r
1828 ldmfd sp!, {r4-r10,lr}
\r
1831 #endif /* UNALIGNED_DRAWLINEDEST */
\r
1834 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
\r
1837 .global blockcpy @ void *dst, void *src, size_t n
\r
1840 stmfd sp!, {r4,r5}
\r
1841 mov r2, r2, lsr #4
\r
1843 ldmia r1!, {r3-r5,r12}
\r
1845 stmia r0!, {r3-r5,r12}
\r
1847 ldmfd sp!, {r4,r5}
\r
1851 .global blockcpy_or @ void *dst, void *src, size_t n, int pat
\r
1854 stmfd sp!, {r4-r6}
\r
1855 orr r3, r3, r3, lsl #8
\r
1856 orr r3, r3, r3, lsl #16
\r
1857 mov r2, r2, lsr #4
\r
1859 ldmia r1!, {r4-r6,r12}
\r
1865 stmia r0!, {r4-r6,r12}
\r
1866 bne blockcpy_loop_or
\r
1867 ldmfd sp!, {r4-r6}
\r
1870 @ vim:filetype=armasm
\r