X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=pico%2Fdraw_arm.S;h=d53e04bde73e9c59cf20995af4991128d8385200;hb=88fd63ad10faa746ef9d7ad7d98a72e51fe2aa86;hp=2c1db10424810425db98000ad20a2d98a66cfa2e;hpb=ea38612fad50103e224a3d00492d40b7dcff9e94;p=picodrive.git diff --git a/pico/draw_arm.S b/pico/draw_arm.S index 2c1db10..d53e04b 100644 --- a/pico/draw_arm.S +++ b/pico/draw_arm.S @@ -10,15 +10,7 @@ #include "pico_int_o32.h" -.extern PicoOpt -.extern HighCol -.extern HighSprZ -.extern HighPreSpr -.extern DrawLineDest .extern DrawStripInterlace -.extern HighCacheS_ptr - -.equiv OVERRIDE_HIGHCOL, 1 .equ PDRAW_SPRITES_MOVED, (1<<0) .equ PDRAW_WND_DIFF_PRIO, (1<<1) @@ -286,7 +278,7 @@ @ int cells; // 0x14 @ }; -@ void DrawLayer(int plane_sh, int *hcache, int cellskip, int maxcells, +@ void DrawLayer(int lflags, int *hcache, int cellskip, int maxcells, @ struct PicoEState *est) .global DrawLayer @@ -295,14 +287,14 @@ DrawLayer: ldr r12, [sp] @ est stmfd sp!, {r4-r11,lr} - ldr r11, [r12, #OFS_Pico_video] + ldr r11, [r12, #OFS_EST_Pico] mov r8, #1 - ldrb r7, [r11, #16] @ ??vv??hh + ldrb r7, [r11, #OFS_Pico_video_reg+16] @ ??vv??hh mov r6, r1 @ hcache - orr r9, r3, r0, lsl #30 - orr r9, r9, r2, lsl #8 @ r9=sh[31]|cellskip[15:8]|maxcells[7:0] (tmp) + orr r9, r3, r0, lsl #29 @ r9=force[31]|sh[30]|plane[29] + orr r9, r9, r2, lsl #8 @ |cellskip[15:8]|maxcells[7:0] (tmp) mov r1, r7, lsl #4 orr r1, r1, #0x00ff @@ -311,19 +303,21 @@ DrawLayer: cmp r10, #1 biclt r1, r1, #0xfc00 biceq r1, r1, #0xfe00 - bicgt r1, r1, #0xff00 @ r1=ymask=(height<<8)|0xff; ...; // Y Mask in pixels + cmp r10, #2 + moveq r1, #0x0007 + movgt r1, #0x00ff @ r1=ymask=(height<<8)|0xff; ...; // Y Mask in pixels add r10, r10, #5 cmp r10, #7 subge r10, r10, #1 @ r10=shift[width] (5,6,6,7) - ldr r2, [r12, #OFS_DrawScanline] - ldr lr, [r12, #OFS_Pico_vram] + ldr r2, [r12, #OFS_EST_DrawScanline] + ldr lr, [r12, #OFS_EST_PicoMem_vram] @ Find name table: ands r0, r0, #1 - ldreqb r12, [r11, #2] - ldrneb r12, [r11, #4] + ldreqb r12, [r11, #OFS_Pico_video_reg+2] + ldrneb r12, [r11, #OFS_Pico_video_reg+4] @ calculate xmask: mov r5, r8, lsl r10 @@ -333,8 +327,8 @@ DrawLayer: movne r12, r12, lsl #13 and r12, r12, #(7<<13) @ r12=(ts->nametab<<1) (halfword compliant) - ldrh r8, [r11, #12] - ldrb r7, [r11, #11] + ldrh r8, [r11, #OFS_Pico_video_reg+12] + ldrb r7, [r11, #OFS_Pico_video_reg+11] mov r4, r8, lsr #8 @ pvid->reg[13] mov r4, r4, lsl #10 @ htab=pvid->reg[13]<<9; (halfwords) @@ -351,7 +345,7 @@ DrawLayer: @ Get vertical scroll value: add r7, lr, #0x012000 - add r7, r7, #0x000180 @ r7=Pico.vsram (Pico+0x22180) + add r7, r7, #0x000180 @ r7=PicoMem.vsram (PicoMem+0x22180) ldr r7, [r7] tst r8, #2 @@ -381,10 +375,11 @@ DrawLayer: and r1, r1, #7 add r7, r1, #1 @ r7=dx=((ts->hscroll-1)&7)+1 - tst r9, #1<<31 + movs r3, r9, lsl #1 @ (force[31]|sh[30]) << 1 mov r3, #0 - orrne r10,r10, #1<<23 @ r10=(cells<<24|sh<<23|hi_not_empty<<22|had_output<<21|ty) - movne r3, #0x40 @ default to shadowed pal on sh mode + orrmi r10,r10, #1<<23 @ r10=cells[31:24]|sh[23]|hi_not_empty[22] + orrcs r10,r10, #1<<20 @ |had_output[21]|force[20]|ty[15:0] + movmi r3, #0x40 @ default to shadowed pal on sh mode cmp r7, #8 addne r10,r10, #0x01000000 @ we will loop cells+1 times if there is scroll @@ -395,17 +390,12 @@ DrawLayer: sub r10,r10,r9, lsl #16 @ cells-=cellskip @ cache some stuff to avoid mem access -.if OVERRIDE_HIGHCOL - ldr r11,=HighCol - mov r0, #0xf - ldr r11,[r11] -.else - ldr r11,=HighCol + ldr r11,[sp, #9*4] @ est mov r0, #0xf -.endif + ldr r11,[r11, #OFS_EST_HighCol] mvn r9, #0 @ r9=prevcode=-1 - add r1, r11, r7 @ r1=pdest + add r1, r11, r7 @ r1=pdest @ r4 & r7 are scratch in this loop @@ -424,6 +414,7 @@ DrawLayer: add r8, r8, #1 tst r7, #0x8000 + tsteq r10, #1<<20 @ force? bne .DrawStrip_hiprio cmp r7, r9 @@ -491,7 +482,7 @@ DrawLayer: orr r10, r10, #1<<22 .DrawStrip_hiprio: - tst r10, #0x00c00000 + tst r10, #0x00d00000 @ sh[23]|hi_not_empty[22]|force[20] beq .DrawStrip_hiprio_maybempt sub r0, r1, r11 orr r7, r7, r0, lsl #16 @@ -506,10 +497,10 @@ DrawLayer: tst r10, #1<<21 @ seen non hi-prio tile ldr r1, [sp, #9*4] @ est mov r0, #0 - ldreq r2, [r1, #OFS_rendstatus] + ldreq r2, [r1, #OFS_EST_rendstatus] str r0, [r6] @ terminate the cache list orreq r2, r2, #PDRAW_PLANE_HI_PRIO @ had a layer with all hi-prio tiles - streq r2, [r1, #OFS_rendstatus] + streq r2, [r1, #OFS_EST_rendstatus] ldmfd sp!, {r4-r11,lr} bx lr @@ -524,7 +515,7 @@ DrawLayer: ldr r11, [sp, #9*4] @ est orr r5, r1, r10, lsl #24 - ldr r4, [r11, #OFS_DrawScanline] + ldr r4, [r11, #OFS_EST_DrawScanline] sub r1, r3, #1 orr r5, r5, r4, lsl #16 @ r5=(shift_width[31:24]|scanline[23:16]|ymask[15:0]) and r1, r1, #7 @@ -548,14 +539,9 @@ DrawLayer: add r10,r10,r9, lsl #16 @ cell+=cellskip @ cache some stuff to avoid mem access -.if OVERRIDE_HIGHCOL - ldr r11,=HighCol + ldr r11,[sp, #9*4] @ est mov r0, #0xf - ldr r11,[r11] -.else - ldr r11,=HighCol - mov r0, #0xf -.endif + ldr r11,[r11, #OFS_EST_HighCol] mvn r9, #0 @ r9=prevcode=-1 add r1, r11, r7 @ r1=pdest @@ -571,7 +557,7 @@ DrawLayer: @ calc offset and read tileline code to r7, also calc ty add r7, lr, #0x012000 - add r7, r7, #0x000180 @ r7=Pico.vsram (Pico+0x22180) + add r7, r7, #0x000180 @ r7=PicoMem.vsram (PicoMem+0x22180) add r7, r7, r10,asr #23 @ vsram + ((cell&~1)<<1) bic r7, r7, #3 tst r10,#0x8000 @ plane1? @@ -590,7 +576,7 @@ DrawLayer: mov r4, r4, lsl r7 @ nametabadd and r7, r8, r8, lsr #25 - add r7, lr, r7, lsl #1 @ Pico.vram+((tilex&ts->xmask) as halfwords) + add r7, lr, r7, lsl #1 @ PicoMem.vram+((tilex&ts->xmask) as halfwords) add r7, r7, r4, lsl #1 ldrh r7, [r7, r12] @ r7=code (int, but from unsigned, no sign extend) @@ -612,7 +598,7 @@ DrawLayer: mov r2, r2, lsr #17 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe; - ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels + ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(PicoMem.vram+addr); // Get 8 pixels bic r7, r3, #0x3f and r3, r9, #0x6000 @@ -670,7 +656,7 @@ DrawLayer: add r2, r2, r10, lsl #17 mov r2, r2, lsr #17 eorcs r2, r2, #0x0e @ if (code&0x1000) addr^=0xe; - ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels + ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(PicoMem.vram+addr); // Get 8 pixels mov r9, r7 @ remember code tst r2, r2 orrne r10, r10, #1<<22 @@ -681,10 +667,10 @@ DrawLayer: tst r8, #(1<<24) @ seen non hi-prio tile ldr r1, [sp, #9*4] @ est mov r0, #0 - ldreq r2, [r1, #OFS_rendstatus] + ldreq r2, [r1, #OFS_EST_rendstatus] str r0, [r6] @ terminate the cache list orreq r2, r2, #PDRAW_PLANE_HI_PRIO @ had a layer with all hi-prio tiles - streq r2, [r1, #OFS_rendstatus] + streq r2, [r1, #OFS_EST_rendstatus] ldmfd sp!, {r4-r11,lr} bx lr @@ -722,23 +708,17 @@ DrawLayer: @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ +@ void BackFill(int reg7, int sh, struct PicoEState *est) -.global BackFill @ int reg7, int sh +.global BackFill BackFill: stmfd sp!, {r4-r9,lr} -.if OVERRIDE_HIGHCOL - ldr lr, =HighCol mov r0, r0, lsl #26 - ldr lr, [lr] + ldr lr, [r2, #OFS_EST_HighCol] mov r0, r0, lsr #26 add lr, lr, #8 -.else - ldr lr, =(HighCol+8) - mov r0, r0, lsl #26 - mov r0, r0, lsr #26 -.endif orr r0, r0, r1, lsl #6 orr r0, r0, r0, lsl #8 @@ -764,8 +744,8 @@ BackFill: stmia lr!, {r0-r7} stmia lr!, {r0-r7} - ldmfd sp!, {r4-r9,r12} - bx r12 + ldmfd sp!, {r4-r9,lr} + bx lr @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ @@ -778,15 +758,9 @@ DrawTilesFromCache: stmfd sp!, {r4-r9,r11,lr} @ cache some stuff to avoid mem access -.if OVERRIDE_HIGHCOL - ldr r11,=HighCol + ldr r11,[r3, #OFS_EST_HighCol] mov r12,#0xf - ldr r11,[r11] -.else - ldr r11,=HighCol - mov r12,#0xf -.endif - ldr lr, [r3, #OFS_Pico_vram] + ldr lr, [r3, #OFS_EST_PicoMem_vram] mov r9, r3 @ est mvn r5, #0 @ r5=prevcode=-1 @@ -802,7 +776,8 @@ DrawTilesFromCache: bic r4, r1, #0xfe00 add r1, r11, r4 @ r1=pdest - mov r7, r6, lsl #16 + movs r7, r6, lsl #16 + bpl .dtfc_loop @ !(code & 0x8000) cmp r5, r7, lsr #16 beq .dtfc_samecode @ if (code==prevcode) @@ -917,14 +892,14 @@ DrawTilesFromCache: @ check if we have detected layer covered with hi-prio tiles: .dtfc_check_rendflags: - ldr r2, [r9, #OFS_rendstatus] + ldr r2, [r9, #OFS_EST_rendstatus] tst r2, #(PDRAW_PLANE_HI_PRIO|PDRAW_SHHI_DONE) beq .dtfc_loop bic r8, r8, #1 @ sh/hi mode off tst r2, #PDRAW_SHHI_DONE bne .dtfc_loop @ already processed orr r2, r2, #PDRAW_SHHI_DONE - str r2, [r9, #OFS_rendstatus] + str r2, [r9, #OFS_EST_rendstatus] add r1, r11,#8 mov r3, #320/4/4 @@ -964,22 +939,16 @@ DrawSpritesSHi: add r10,r0, #3 @ r10=HighLnSpr end add r10,r10,r3 @ r10=HighLnSpr end -.if OVERRIDE_HIGHCOL - ldr r11,=HighCol + ldr r11,[r1, #OFS_EST_HighCol] mov r12,#0xf - ldr r11,[r11] -.else - ldr r11,=HighCol - mov r12,#0xf -.endif - ldr lr, [r1, #OFS_Pico_vram] + ldr lr, [r1, #OFS_EST_PicoMem_vram] DrawSpriteSHi: @ draw next sprite ldrb r0, [r10,#-1]! ldr r7, [sp] @ est - ldr r1, =HighPreSpr + ldr r1, [r7, #OFS_EST_HighPreSpr] cmp r0, #0xff ldmeqfd sp!, {r1,r4-r11,pc} @ end of list and r0, r0, #0x7f @@ -1005,7 +974,7 @@ DrawSpriteSHi: mov r5, r3, lsr #24 and r5, r5, #7 @ r5=height - ldr r7, [r7, #OFS_DrawScanline] + ldr r7, [r7, #OFS_EST_DrawScanline] mov r0, r3, lsl #16 @ r4=sy<<16 (tmp) sub r7, r7, r0, asr #16 @ r7=row=DrawScanline-sy @@ -1046,7 +1015,7 @@ DrawSpriteSHi: mov r8, r8, lsl #17 mov r8, r8, lsr #17 @ tile&=0x7fff; // Clip tile address - ldr r2, [lr, r8, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels + ldr r2, [lr, r8, lsl #1] @ pack=*(unsigned int *)(PicoMem.vram+addr); // Get 8 pixels add r1, r11, r0 @ r1=pdest tst r2, r2 beq .dsprShi_loop @@ -1137,17 +1106,6 @@ DrawSpriteSHi: DrawAllSprites: orr r1, r2, r1, lsl #1 - ldr r12,[r3, #OFS_rendstatus] - tst r12,#(PDRAW_DIRTY_SPRITES|PDRAW_SPRITES_MOVED) - beq das_no_prep - stmfd sp!, {r0,r1,r3,lr} - and r0, r12,#PDRAW_DIRTY_SPRITES - bic r12,r12,#(PDRAW_DIRTY_SPRITES|PDRAW_SPRITES_MOVED) - str r12,[r3, #OFS_rendstatus] - bl PrepareSprites - ldmfd sp!, {r0,r1,r3,lr} - -das_no_prep: ldr r2, [r0] ands r2, r2, #0x7f bxeq lr @@ -1159,15 +1117,9 @@ das_no_prep: add r10,r0, #3 add r10,r10,r2 @ r10=HighLnSpr end -.if OVERRIDE_HIGHCOL - ldr r11,=HighCol + ldr r11,[r3, #OFS_EST_HighCol] mov r12,#0xf - ldr r11,[r11] -.else - ldr r11,=HighCol - mov r12,#0xf -.endif - ldr lr, [r3, #OFS_Pico_vram] + ldr lr, [r3, #OFS_EST_PicoMem_vram] @ + 0 : hhhhvvvv ----hhvv yyyyyyyy yyyyyyyy // v, h: horiz. size @ + 4 : xxxxxxxx xxxxxxxx pccvhnnn nnnnnnnn // x: x coord + 8 @@ -1182,12 +1134,12 @@ DrawSprite: ldmeqfd sp!, {r1,r3-r11,pc} @ end of list cmp r2, r8, lsr #1 bne DrawSprite @ wrong priority - ldr r1, =HighPreSpr + ldr r1, [r7, #OFS_EST_HighPreSpr] and r0, r0, #0x7f add r0, r1, r0, lsl #3 ldr r3, [r0] @ sprite[0] - ldr r7, [r7, #OFS_DrawScanline] + ldr r7, [r7, #OFS_EST_DrawScanline] mov r6, r3, lsr #28 sub r6, r6, #1 @ r6=width-1 (inc later) mov r5, r3, lsr #24 @@ -1246,7 +1198,7 @@ DrawSprite: mov r8, r8, lsl #17 mov r8, r8, lsr #17 @ tile&=0x7fff; // Clip tile address - ldr r2, [lr, r8, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels + ldr r2, [lr, r8, lsl #1] @ pack=*(unsigned int *)(PicoMem.vram+addr); // Get 8 pixels add r1, r11, r0 @ r1=pdest tst r2, r2 beq .dspr_loop @@ -1315,12 +1267,12 @@ DrawWindow: ldr r12, [sp] @ est stmfd sp!, {r4-r11,lr} - ldr r6, [r12, #OFS_Pico_video] - ldr r10, [r12, #OFS_DrawScanline] + ldr r6, [r12, #OFS_EST_Pico] + ldr r10, [r12, #OFS_EST_DrawScanline] mov r11, r12 @ est - ldrb r12, [r6, #3] @ pvid->reg[3] + ldrb r12, [r6, #OFS_Pico_video_reg+3] @ pvid->reg[3] - ldr r4, [r6, #12] + ldr r4, [r6, #OFS_Pico_video_reg+12] mov r5, r10, lsr #3 and r10, r10, #7 mov r10, r10, lsl #1 @ r10=ty @@ -1334,8 +1286,8 @@ DrawWindow: addeq r12, r12, r5, lsl #6 @ nametab add r12, r12, r0, lsl #2 @ +starttile - ldr lr, [r11, #OFS_Pico_vram] - ldr r6, [r11, #OFS_rendstatus] + ldr lr, [r11, #OFS_EST_PicoMem_vram] + ldr r6, [r11, #OFS_EST_rendstatus] @ fetch the first code now ldrh r7, [lr, r12] @@ -1352,17 +1304,10 @@ DrawWindow: sub r8, r1, r0 @ cache some stuff to avoid mem access -.if OVERRIDE_HIGHCOL - ldr r11,=HighCol + ldr r11, [r11, #OFS_EST_HighCol] mov r8, r8, lsl #1 @ cells - ldr r11,[r11] - mvn r9, #0 @ r9=prevcode=-1 add r11,r11,#8 -.else - ldr r11,=(HighCol+8) - mov r8, r8, lsl #1 @ cells mvn r9, #0 @ r9=prevcode=-1 -.endif add r1, r11, r0, lsl #4 @ r1=pdest mov r0, #0xf b .dwloop_enter @@ -1447,9 +1392,9 @@ DrawWindow: and r2, r6, #PDRAW_WND_DIFF_PRIO ldmfd sp!, {r4-r11,lr} ldr r0, [sp] - ldr r1, [r0, #OFS_rendstatus] + ldr r1, [r0, #OFS_EST_rendstatus] orr r1, r1, r2 - str r1, [r0, #OFS_rendstatus] + str r1, [r0, #OFS_EST_rendstatus] bx lr @@ -1541,17 +1486,17 @@ PicoDoHighPal555: stmfd sp!, {r4-r10,lr} mov r10,r2 @ est mov r1, #0 - ldr r8, [r10, #OFS_Pico_video] + ldr r8, [r10, #OFS_EST_Pico] PicoDoHighPal555_nopush: orr r9, r1, r0, lsl #31 @ 0:called from FinalizeLine555, 31: s/h - ldr r0, =HighPal + add r0, r10, #OFS_EST_HighPal mov r1, #0 - strb r1, [r8, #-0x1a] @ 0x2220e ~ dirtyPal + strb r1, [r8, #OFS_Pico_m_dirtyPal] - sub r1, r8, #0x128 @ r1=Pico.cram + ldr r1, [r10, #OFS_EST_PicoMem_cram] mov r2, #0x40 mov r8, #0x0061 orr r8, r8, #0x0800 @@ -1562,7 +1507,7 @@ PicoDoHighPal555_nopush: tst r9, #(1<<31) beq PicoDoHighPal555_end - ldr r3, =HighPal + add r3, r10, #OFS_EST_HighPal @ shadowed pixels: mov r12, #0x008e @@ -1605,7 +1550,7 @@ PicoDoHighPal555_end: tst r9, #1 ldmeqfd sp!, {r4-r10,pc} - ldr r8, [r10, #OFS_Pico_video] + ldr r8, [r10, #OFS_EST_Pico] b FinalizeLineRGB555_pal_done @@ -1616,41 +1561,33 @@ PicoDoHighPal555_end: FinalizeLine555: stmfd sp!, {r4-r10,lr} mov r10,r2 @ est - ldr r8, [r10, #OFS_Pico_video] + ldr r8, [r10, #OFS_EST_Pico] - ldrb r2, [r8, #-0x1a] @ 0x2220e ~ dirtyPal + ldrb r2, [r8, #OFS_Pico_m_dirtyPal] mov r1, #1 tst r2, r2 bne PicoDoHighPal555_nopush FinalizeLineRGB555_pal_done: - ldr r3, =HighPal + add r3, r10, #OFS_EST_HighPal - ldr r12, [r10, #OFS_rendstatus] + ldr r12, [r10, #OFS_EST_rendstatus] eors r0, r0, #1 @ sh is 0 mov lr, #0xff tstne r12,#PDRAW_ACC_SPRITES movne lr, #0x3f -.if OVERRIDE_HIGHCOL - ldr r1, =HighCol - ldr r0, =DrawLineDest - ldr r1, [r1] - ldr r0, [r0] + ldr r1, [r10, #OFS_EST_HighCol] + ldr r0, [r10, #OFS_EST_DrawLineDest] add r1, r1, #8 -.else - ldr r0, =DrawLineDest - ldr r1, =(HighCol+8) - ldr r0, [r0] -.endif - ldrb r12, [r8, #12] + ldrb r12, [r8, #OFS_Pico_video_reg+12] mov lr, lr, lsl #1 tst r12, #1 movne r2, #320/8 @ len bne .fl_no32colRGB555 - ldr r4, =PicoOpt + ldr r4, [r10, #OFS_EST_PicoOpt] mov r2, #256/8 ldr r4, [r4] tst r4, #0x4000