asm for DrawStripVSRam

git-svn-id: file:///home/notaz/opt/svn/PicoDrive@223 be3aeb3a-fb24-0410-a615-afba39da0efa

diff --git a/Pico/Draw.c b/Pico/Draw.c
index 2eed289..582c8cb 100644 (file)
--- a/Pico/Draw.c
+++ b/Pico/Draw.c
@@ -34,7 +34,7 @@ static int SpriteBlocks;
 struct TileStrip\r
 {\r
   int nametab; // Position in VRAM of name table (for this tile line)\r
-  int line;    // Line number in pixels 0x000-0x3ff within the virtual tilemap \r
+  int line;    // Line number in pixels 0x000-0x3ff within the virtual tilemap\r
   int hscroll; // Horizontal scroll value in pixels for the line\r
   int xmask;   // X-Mask (0x1f - 0x7f) for horizontal wraparound in the tilemap\r
   int *hc;     // cache for high tile codes and their positions\r
@@ -317,12 +317,8 @@ static void DrawStrip(struct TileStrip *ts, int sh)
   // terminate the cache list\r
   *ts->hc = 0;\r
 }\r
-#endif\r
 \r
 // this is messy\r
-#ifndef _ASM_DRAW_C\r
-static\r
-#endif\r
 void DrawStripVSRam(struct TileStrip *ts, int plane)\r
 {\r
   int tilex=0,dx=0,ty=0,code=0,addr=0,cell=0,nametabadd=0;\r
@@ -386,6 +382,7 @@ void DrawStripVSRam(struct TileStrip *ts, int plane)
   // terminate the cache list\r
   *ts->hc = 0;\r
 }\r
+#endif\r
 \r
 #ifndef _ASM_DRAW_C\r
 static\r
@@ -921,13 +918,13 @@ static void PrepareSprites(int full)
       height = (hv&3)+1;\r
 \r
       if(sy > 240 || sy + (height<<3) <= 0) skip|=1<<22;\r
-      \r
+\r
       width  = (hv>>2)+1;\r
       code2 = sprite[1];\r
       sx = (code2>>16)&0x1ff;\r
       sx -= 0x78; // Get X coordinate + 8\r
       sx_min = 8-(width<<3);\r
-    \r
+\r
       if((sx <= sx_min && sx >= -0x76) || sx >= 328) skip|=1<<23;\r
       else if (sx > sx_min && !skip) {\r
         int sbl = (2<<height)-1;\r
@@ -935,10 +932,10 @@ static void PrepareSprites(int full)
         if(shi < 0) shi=0; // negative sy\r
         sblocks |= sbl<<shi;\r
       }\r
-    \r
+\r
       *pd++ = (width<<28)|(height<<24)|skip|(hv<<16)|((unsigned short)sy);\r
       *pd++ = (sx<<16)|((unsigned short)code2);\r
-    \r
+\r
       // Find next sprite\r
       link=(code>>16)&0x7f;\r
       if(!link) break; // End of sprites\r
@@ -1205,7 +1202,7 @@ static int DrawDisplay(int sh)
   if (win&0x80) { if (Scanline>=edge) hvwind=1; }\r
   else          { if (Scanline< edge) hvwind=1; }\r
 \r
-  if(!hvwind) { // we might have a vertical window here \r
+  if(!hvwind) { // we might have a vertical window here\r
     win=pvid->reg[0x11];\r
     edge=win&0x1f;\r
     if(win&0x80) {\r

diff --git a/Pico/Draw.s b/Pico/Draw.s
index 7643c3d..cb1a6fa 100644 (file)
--- a/Pico/Draw.s
+++ b/Pico/Draw.s
@@ -14,7 +14,6 @@
 .extern HighSprZ\r
 .extern rendstatus\r
 .extern DrawLineDest\r
-.extern DrawStripVSRam\r
 .extern DrawStripInterlace\r
 \r
 \r
@@ -293,7 +292,6 @@ DrawLayer:
     add     r12, r12, r4, lsl r10  @ nametab+=(ts.line>>3)<<shift[width];\r
 \r
     @ ldmia   r0, {r1,r2,r3,r5,r6,r9} @ r2=line, r3=ts->hscroll, r5=ts->xmask, r6=ts->hc, r9=ts->cells\r
-@    mov     r12,r1,  lsl #1 @ r12=(ts->nametab<<1) (halfword compliant)\r
 \r
     and     r10,r2,  #7\r
     mov     r10,r10, lsl #1 @ r10=ty=(ts->line&7)<<1;\r
@@ -422,31 +420,159 @@ DrawLayer:
     ldmfd   sp!, {r4-r11,lr}\r
     bx      lr\r
 \r
+@ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@\r
 \r
 .DrawStrip_vsscroll:\r
-    @ shit, we have 2-cell column based vscroll\r
-    @ let the c code handle this (for now)\r
+    rsb     r8, r3, #0\r
+    mov     r8, r8, lsr #3        @ r8=tilex=(-ts->hscroll)>>3\r
+    bic     r8, r8, #0xff000000\r
+    orr     r8, r8, r5, lsl #25   @ r8=(xmask[31:25]|tilex[15:0])\r
 \r
-    @   int nametab; // 0x00\r
-    @   int line;    // 0x04\r
-    @   int hscroll; // 0x08\r
-    @   int xmask;   // 0x0C\r
-    @   int *hc;     // 0x10 (pointer to cache buffer)\r
-    @   int cells;   // 0x14\r
+    ldr     r4, =Scanline\r
+    orr     r5, r1, r10, lsl #24\r
+    ldr     r4, [r4]\r
+    sub     r1, r3, #1\r
+    orr     r5, r5, r4, lsl #16   @ r5=(shift_width[31:24]|scanline[23:16]|ymask[15:0])\r
+    and     r1, r1, #7\r
+    add     r7, r1, #1            @ r7=dx=((ts->hscroll-1)&7)+1\r
 \r
-    sub     sp, sp, #6*4\r
-    orr     r2, r1, r10, lsl #24 @ ts.line=ymask|(shift[width]<<24); // save some stuff instead of line\r
-    mov     r1, r0               @ plane\r
-    mov     r0, r12, lsr #1      @ halfwords\r
-    and     r9, r9, #0xff\r
-    stmia   sp, {r0,r2,r3,r5,r6,r9}\r
+    mov     r10,r9, lsl #16\r
+    tst     r0, r0\r
+    orrne   r10,r10, #0x8000\r
+    tst     r9, #1<<31\r
+    mov     r3, #0\r
+    orr     r10,r10, #0xff000000 @ will be adjusted on entering loop\r
+    orrne   r10,r10, #1<<23 @ r10=(cells[31:24]|sh[23]|hi_not_empty[22]|cells_max[21:16]|plane[15]|ty[14:0])\r
+    movne   r3, #0x40       @ default to shadowed pal on sh mode\r
 \r
-    mov     r0, sp\r
-    bl      DrawStripVSRam @ struct TileStrip *ts, int plane\r
+    mvn     r9, #0          @ r9=prevcode=-1\r
 \r
-    add     sp, sp, #6*4\r
-    ldmfd   sp!, {r4-r11,lr}\r
-    bx      lr\r
+    @ cache some stuff to avoid mem access\r
+    ldr     r11,=HighCol\r
+    mov     r0, #0xf\r
+    add     r1, r11, r7         @ r1=pdest\r
+\r
+    cmp     r7, #8\r
+    subne   r10,r10, #0x01000000 @ have hscroll, start with negative cell\r
+\r
+\r
+    @ r4 & r7 are scratch in this loop\r
+.dsloop_vs_subr1:\r
+    sub     r1, r1, #8\r
+.dsloop_vs: @ 40-41 times\r
+    add     r10,r10, #0x01000000\r
+    and     r4, r10, #0x003f0000\r
+    cmp     r4, r10, asr #8\r
+    ble     .dsloop_exit\r
+\r
+    @ calc offset and read tileline code to r7, also calc ty\r
+    add     r7, lr, #0x012000\r
+    add     r7, r7, #0x000180     @ r7=Pico.vsram (Pico+0x22180)\r
+    add     r7, r7, r10,asr #23   @ vsram + ((cell&~1)<<1)\r
+    bic     r7, r7, #3\r
+    tst     r10,#0x8000           @ plane1?\r
+    addne   r7, r7, #2\r
+    ldrh    r7, [r7]              @ r7=vscroll\r
+\r
+    bic     r10,r10,#0xff         @ clear old ty\r
+    and     r4, r5, #0xff0000\r
+    add     r4, r4, r7, lsl #16\r
+    and     r4, r4, r5, lsl #16   @ r4=line<<16\r
+    and     r7, r4, #0x70000\r
+    orr     r10,r10,r7, lsr #15   @ new ty\r
+\r
+    mov     r4, r4, lsr #19\r
+    mov     r7, r5, lsr #24\r
+    mov     r4, r4, lsl r7        @ nametabadd\r
+\r
+    and     r7, r8, r8, lsr #25\r
+    add     r7, lr, r7, lsl #1    @ Pico.vram+((tilex&ts->xmask) as halfwords)\r
+    add     r7, r7, r4, lsl #1\r
+    ldrh    r7, [r7, r12]         @ r7=code (int, but from unsigned, no sign extend)\r
+\r
+    add     r1, r1, #8\r
+    add     r8, r8, #1\r
+\r
+    tst     r7, #0x8000\r
+    bne     .DrawStrip_vs_hiprio\r
+\r
+    cmp     r7, r9\r
+    beq     .DrawStrip_vs_samecode @ we know stuff about this tile already\r
+\r
+    mov     r9, r7          @ remember code\r
+\r
+    movs    r2, r9, lsl #20 @ if (code&0x1000)\r
+    mov     r2, r2, lsl #1\r
+    add     r2, r2, r10, lsl #17\r
+    mov     r2, r2, lsr #17\r
+    eorcs   r2, r2, #0x0e   @ if (code&0x1000) addr^=0xe;\r
+\r
+    ldr     r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels\r
+\r
+    bic     r7, r3, #0x3f\r
+    and     r3, r9, #0x6000\r
+    add     r3, r7, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);\r
+\r
+.DrawStrip_vs_samecode:\r
+    tst     r2, r2\r
+    beq     .dsloop_vs              @ tileline blank\r
+\r
+    cmp     r2, r2, ror #4\r
+    beq     .DrawStrip_vs_SingleColor @ tileline singlecolor \r
+\r
+    tst     r9, #0x0800\r
+    beq     .DrawStrip_vs_TileNorm\r
+\r
+    @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r0: helper pattern\r
+    TileFlip r0\r
+    b       .dsloop_vs\r
+\r
+.DrawStrip_vs_TileNorm:\r
+    TileNorm r0\r
+    b       .dsloop_vs\r
+\r
+.DrawStrip_vs_SingleColor:\r
+    and     r4, r2, #0xf\r
+    orr     r4, r3, r4\r
+    orr     r4, r4, r4, lsl #8\r
+    tst     r1, #1             @ not aligned?\r
+    strneb  r4, [r1], #1\r
+    streqh  r4, [r1], #2\r
+    strh    r4, [r1], #2\r
+    strh    r4, [r1], #2\r
+    strh    r4, [r1], #2\r
+    strneb  r4, [r1], #1       @ have a remaining unaligned pixel?\r
+    b       .dsloop_vs_subr1\r
+\r
+.DrawStrip_vs_hiprio:\r
+    tst     r10, #0x00c00000\r
+    beq     .DrawStrip_vs_hiprio_maybempt\r
+    sub     r0, r1, r11\r
+    orr     r7, r7, r0,  lsl #16\r
+    orr     r7, r7, r10, lsl #25 @ (ty<<25)\r
+    tst     r7, #0x1000\r
+    eorne   r7, r7, #7<<26  @ if(code&0x1000) cval^=7<<26;\r
+    str     r7, [r6], #4    @ cache hi priority tile\r
+    mov     r0, #0xf\r
+    b       .dsloop_vs\r
+\r
+.DrawStrip_vs_hiprio_maybempt:\r
+    cmp     r7, r9\r
+    beq     .dsloop_vs         @ must've been empty, otherwise we wouldn't get here\r
+    movs    r2, r7, lsl #20 @ if (code&0x1000)\r
+    mov     r2, r2, lsl #1\r
+    add     r2, r2, r10, lsl #17\r
+    mov     r2, r2, lsr #17\r
+    eorcs   r2, r2, #0x0e   @ if (code&0x1000) addr^=0xe;\r
+    ldr     r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels\r
+    mov     r9, r7          @ remember code\r
+    tst     r2, r2\r
+    orrne   r10, r10, #1<<22\r
+    bne     .DrawStrip_vs_hiprio\r
+    b       .dsloop_vs\r
+\r
+\r
+@ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@\r
 \r
 @ interlace mode 2? Sonic 2?\r
 .DrawStrip_interlace:\r