vdp rendering, fixes and optimisations

diff --git a/pico/32x/draw.c b/pico/32x/draw.c
index 991abae..bef8dd3 100644 (file)
--- a/pico/32x/draw.c
+++ b/pico/32x/draw.c
@@ -31,7 +31,11 @@ static void convert_pal555(int invert_prio)
   // place prio to LS green bit
   for (i = 0x100/2; i > 0; i--, ps++, pd++) {
     unsigned int t = *ps;
+#if defined(USE_BGR555)
+    *pd = t ^ inv;
+#else
     *pd = (((t & m1) << 11) | ((t & m2) << 1) | ((t & m3) >> 10)) ^ inv;
+#endif
   }
 
   Pico32x.dirty_pal = 0;

diff --git a/pico/draw.c b/pico/draw.c
index 02dffec..a57523a 100644 (file)
--- a/pico/draw.c
+++ b/pico/draw.c
@@ -56,8 +56,8 @@ static unsigned int DefOutBuff[320*2/2];
 void *DrawLineDestBase = DefOutBuff;\r
 int DrawLineDestIncrement;\r
 \r
-static int  HighCacheA[41+1];   // caches for high layers\r
-static int  HighCacheB[41+1];\r
+static int  HighCacheA[41*2+1]; // caches for high layers\r
+static int  HighCacheB[41*2+1];\r
 static int  HighPreSpr[80*2+1]; // slightly preprocessed sprites\r
 \r
 unsigned int VdpSATCache[128];  // VDP sprite cache (1st 32 sprite attr bits)\r
@@ -275,6 +275,7 @@ static void DrawStrip(struct TileStrip *ts, int lflags, int cellskip)
   int tilex,dx,ty,code=0,addr=0,cells;\r
   int oldcode=-1,blank=-1; // The tile we know is blank\r
   int pal=0;\r
+  unsigned int pack = 0;\r
 \r
   // Draw tiles across screen:\r
   tilex=((-ts->hscroll)>>3)+cellskip;\r
@@ -286,33 +287,36 @@ static void DrawStrip(struct TileStrip *ts, int lflags, int cellskip)
 \r
   for (; cells > 0; dx+=8, tilex++, cells--)\r
   {\r
-    unsigned int pack;\r
-\r
     code = PicoMem.vram[ts->nametab + (tilex & ts->xmask)];\r
-    if ((code >> 15) | (lflags & LF_FORCE)) { // high priority tile\r
-      int cval = code | (dx<<16) | (ty<<25);\r
-      if(code&0x1000) cval^=7<<26;\r
-      *ts->hc++ = cval; // cache it\r
-      continue;\r
-    }\r
-    if (code == blank)\r
-      continue;\r
 \r
     if (code!=oldcode) {\r
       oldcode = code;\r
-      // Get tile address/2:\r
-      addr=(code&0x7ff)<<4;\r
-      addr+=ty;\r
-      if (code&0x1000) addr^=0xe; // Y-flip\r
 \r
-      pal=((code>>9)&0x30)|((lflags&LF_SH)<<6); // shadow\r
+      pack = 0;\r
+      if (code != blank) {\r
+        // Get tile address/2:\r
+        addr=(code&0x7ff)<<4;\r
+        addr+=ty;\r
+        if (code&0x1000) addr^=0xe; // Y-flip\r
+\r
+        pal=((code>>9)&0x30)|((lflags&LF_SH)<<6); // shadow\r
+\r
+        pack = *(unsigned int *)(PicoMem.vram + addr);\r
+        if (!pack)\r
+          blank = code;\r
+      }\r
     }\r
 \r
-    pack = *(unsigned int *)(PicoMem.vram + addr);\r
-    if (!pack) {\r
-      blank = code;\r
+    if ((code >> 15) | (lflags & LF_FORCE)) { // high priority tile\r
+      if ((lflags&LF_SH) | (code!=blank)) {\r
+        int cval = code | (dx<<16) | (ty<<25);\r
+        if (code&0x1000) cval^=0xe<<25;\r
+        *ts->hc++ = cval, *ts->hc++ = pack; // cache it\r
+      }\r
       continue;\r
     }\r
+    if (code == blank)\r
+      continue;\r
 \r
     if (code & 0x0800) TileFlip(pd + dx, pack, pal);\r
     else               TileNorm(pd + dx, pack, pal);\r
@@ -362,29 +366,28 @@ static void DrawStripVSRam(struct TileStrip *ts, int plane_sh, int cellskip)
     }\r
 \r
     code=PicoMem.vram[ts->nametab+nametabadd+(tilex&ts->xmask)];\r
-    if (code>>15) { // high priority tile\r
-      int cval = code | (dx<<16) | (ty<<25);\r
-      if(code&0x1000) cval^=7<<26;\r
-      *ts->hc++ = cval; // cache it\r
-      continue;\r
-    }\r
-    if ((code<<16|ty)==blank) continue;\r
-\r
     if (code!=oldcode) {\r
       oldcode = code;\r
       // Get tile address/2:\r
-      addr=(code&0x7ff)<<4;\r
+      addr = (code&0x7ff)<<4;\r
 \r
-      pal=((code>>9)&0x30)|((plane_sh&LF_SH)<<6); // shadow\r
+      pal = ((code>>9)&0x30)|((plane_sh&LF_SH)<<6); // shadow\r
     }\r
+    code |= ty<<16;\r
 \r
     if (code & 0x1000) ty ^= 0xe; // Y-flip\r
     pack = *(unsigned int *)(PicoMem.vram + addr+ty);\r
+    if (!pack)\r
+      blank = code;\r
 \r
-    if (!pack) {\r
-      blank = code<<16|ty;\r
+    if ((code >> 15) | (plane_sh & LF_FORCE)) { // high priority tile\r
+      if ((plane_sh&LF_SH) | (code!=blank)) {\r
+        int cval = code | (dx<<16) | (ty<<25);\r
+        *ts->hc++ = cval, *ts->hc++ = pack; // cache it\r
+      }\r
       continue;\r
     }\r
+    if (code == blank) continue;\r
 \r
     if (code & 0x0800) TileFlip(pd + dx, pack, pal);\r
     else               TileNorm(pd + dx, pack, pal);\r
@@ -405,6 +408,7 @@ void DrawStripInterlace(struct TileStrip *ts, int plane_sh)
   int tilex=0,dx=0,ty=0,code=0,addr=0,cells;\r
   int oldcode=-1,blank=-1; // The tile we know is blank\r
   int pal=0;\r
+  unsigned int pack = 0;\r
 \r
   // Draw tiles across screen:\r
   tilex=(-ts->hscroll)>>3;\r
@@ -415,32 +419,36 @@ void DrawStripInterlace(struct TileStrip *ts, int plane_sh)
 \r
   for (; cells; dx+=8,tilex++,cells--)\r
   {\r
-    unsigned int pack;\r
-\r
     code = PicoMem.vram[ts->nametab + (tilex & ts->xmask)];\r
-    if (code>>15) { // high priority tile\r
-      int cval = (code&0xfc00) | (dx<<16) | (ty<<25);\r
-      cval|=(code&0x3ff)<<1;\r
-      if(code&0x1000) cval^=0xf<<26;\r
-      *ts->hc++ = cval; // cache it\r
-      continue;\r
-    }\r
-    if (code==blank) continue;\r
 \r
     if (code!=oldcode) {\r
       oldcode = code;\r
-      // Get tile address/2:\r
-      addr=(code&0x3ff)<<5;\r
-      if (code&0x1000) addr+=30-ty; else addr+=ty; // Y-flip\r
 \r
-      pal=((code>>9)&0x30)|((plane_sh&LF_SH)<<6); // shadow\r
+      pack = 0;\r
+      if (code != blank) {\r
+        // Get tile address/2:\r
+        addr = (code&0x3ff)<<5;\r
+        if (code&0x1000) addr+=30-ty; else addr+=ty; // Y-flip\r
+\r
+        pal = ((code>>9)&0x30)|((plane_sh&LF_SH)<<6); // shadow\r
+\r
+        pack = *(unsigned int *)(PicoMem.vram + addr);\r
+        if (!pack)\r
+          blank = code;\r
+      }\r
     }\r
 \r
-    pack = *(unsigned int *)(PicoMem.vram + addr);\r
-    if (!pack) {\r
-      blank = code;\r
+    if ((code >> 15) | (plane_sh & LF_FORCE)) { // high priority tile\r
+      if ((plane_sh&LF_SH) | (code!=blank)) {\r
+        int cval = (code&0xfc00) | (dx<<16) | (ty<<25);\r
+        cval|=(code&0x3ff)<<1;\r
+        if (code&0x1000) cval^=0x1e<<25;\r
+        *ts->hc++ = cval, *ts->hc++ = pack; // cache it\r
+      }\r
       continue;\r
     }\r
+    if (code == blank)\r
+      continue;\r
 \r
     if (code & 0x0800) TileFlip(pd + dx, pack, pal);\r
     else               TileNorm(pd + dx, pack, pal);\r
@@ -549,7 +557,7 @@ static void DrawWindow(int tstart, int tend, int prio, int sh,
 \r
   tilex=tstart<<1;\r
 \r
-  if (prio && !(est->rendstatus & PDRAW_WND_HIGH_PRIO)) {\r
+  if (prio && !(est->rendstatus & PDRAW_WND_DIFF_PRIO)) {\r
     // all tiles processed in low prio pass\r
     return;\r
   }\r
@@ -568,7 +576,7 @@ static void DrawWindow(int tstart, int tend, int prio, int sh,
 \r
       code = PicoMem.vram[nametab + tilex];\r
       if ((code>>15) != prio) {\r
-        est->rendstatus |= PDRAW_WND_HIGH_PRIO;\r
+        est->rendstatus |= PDRAW_WND_DIFF_PRIO;\r
         continue;\r
       }\r
       if (code==blank) continue;\r
@@ -600,7 +608,7 @@ static void DrawWindow(int tstart, int tend, int prio, int sh,
 \r
       code = PicoMem.vram[nametab + tilex];\r
       if((code>>15) != prio) {\r
-        est->rendstatus |= PDRAW_WND_HIGH_PRIO;\r
+        est->rendstatus |= PDRAW_WND_DIFF_PRIO;\r
         continue;\r
       }\r
 \r
@@ -651,7 +659,7 @@ static void DrawTilesFromCacheShPrep(void)
 static void DrawTilesFromCache(int *hc, int sh, int rlim, struct PicoEState *est)\r
 {\r
   unsigned char *pd = Pico.est.HighCol;\r
-  int code, addr, dx;\r
+  int code, dx;\r
   unsigned int pack;\r
   int pal;\r
 \r
@@ -667,11 +675,7 @@ static void DrawTilesFromCache(int *hc, int sh, int rlim, struct PicoEState *est
   if (!sh)\r
   {\r
     while ((code=*hc++)) {\r
-      // Get tile address/2:\r
-      addr = (code & 0x7ff) << 4;\r
-      addr += (unsigned int)code >> 25; // y offset into tile\r
-\r
-      pack = *(unsigned int *)(PicoMem.vram + addr);\r
+      pack = *hc++;\r
       if (!pack)\r
         continue;\r
 \r
@@ -689,15 +693,12 @@ static void DrawTilesFromCache(int *hc, int sh, int rlim, struct PicoEState *est
     while ((code=*hc++)) {\r
       unsigned char *zb;\r
 \r
-      // Get tile address/2:\r
-      addr=(code&0x7ff)<<4;\r
-      addr+=(unsigned int)code>>25; // y offset into tile\r
-      dx=(code>>16)&0x1ff;\r
+      dx = (code >> 16) & 0x1ff;\r
       zb = est->HighCol+dx;\r
       *zb++ &= 0x7f; *zb++ &= 0x7f; *zb++ &= 0x7f; *zb++ &= 0x7f;\r
       *zb++ &= 0x7f; *zb++ &= 0x7f; *zb++ &= 0x7f; *zb++ &= 0x7f;\r
 \r
-      pack = *(unsigned int *)(PicoMem.vram + addr);\r
+      pack = *hc++;\r
       if (!pack)\r
         continue;\r
 \r
@@ -880,12 +881,12 @@ static NOINLINE void DrawAllSpritesInterlace(int pri, int sh)
     // check if it is on this line\r
     sy = (code&0x3ff)-0x100;\r
     height = (((code>>24)&3)+1)<<4;\r
-    if(sline < sy || sline >= sy+height) goto nextsprite; // no\r
+    if((sline < sy) | (sline >= sy+height)) goto nextsprite; // no\r
 \r
     // check if sprite is not hidden offscreen\r
     sx = (sx>>16)&0x1ff;\r
     sx -= 0x78; // Get X coordinate + 8\r
-    if(sx <= -8*3 || sx >= 328) goto nextsprite;\r
+    if((sx <= -8*3) | (sx >= 328)) goto nextsprite;\r
 \r
     // sprite is good, save it's pointer\r
     sprites[i++]=sprite;\r
@@ -1494,11 +1495,11 @@ void PicoDoHighPal555_8bit(int sh, int line, struct PicoEState *est)
       t |= (t >> 4) & 0x08610861;\r
       dpal[0x40/2 + i] = t;\r
     }\r
-    // pixels in color 14 always appear normal (hw bug?)\r
+    // shadowed pixels in color 14 always appear normal (hw bug?)\r
     unsigned short *hpal = est->HighPal;\r
-    hpal[0x80 + 0x0e] = hpal[0x40 + 0x0e] = hpal[0x0e];\r
-    hpal[0x80 + 0x1e] = hpal[0x40 + 0x1e] = hpal[0x1e];\r
-    hpal[0x80 + 0x2e] = hpal[0x40 + 0x2e] = hpal[0x2e];\r
+    hpal[0x80 + 0x0e] = hpal[0x0e];\r
+    hpal[0x80 + 0x1e] = hpal[0x1e];\r
+    hpal[0x80 + 0x2e] = hpal[0x2e];\r
   }\r
 }\r
 \r
@@ -1516,7 +1517,7 @@ void PicoDoHighPal555(int sh, int line, struct PicoEState *est)
   for (i = 0; i < 0x40 / 2; i++) {\r
     t = spal[i];\r
 #ifdef USE_BGR555\r
-    t = ((t & 0x0e000e00)<< 3) | ((t & 0x00e000e0)<<2) | ((t & 0x000e000e)<<1);\r
+    t = ((t & 0x000e000e)<< 1) | ((t & 0x00e000e0)<<3) | ((t & 0x0e000e00)<<4);\r
 #else\r
     t = ((t & 0x000e000e)<<12) | ((t & 0x00e000e0)<<3) | ((t & 0x0e000e00)>>7);\r
 #endif\r
@@ -1540,9 +1541,9 @@ void PicoDoHighPal555(int sh, int line, struct PicoEState *est)
     }\r
     // pixels in color 14 always appear normal (hw bug?)\r
     unsigned short *hpal = est->HighPal;\r
-    hpal[0x80 + 0x0e] = hpal[0x40 + 0x0e] = hpal[0x0e];\r
-    hpal[0x80 + 0x1e] = hpal[0x40 + 0x1e] = hpal[0x1e];\r
-    hpal[0x80 + 0x2e] = hpal[0x40 + 0x2e] = hpal[0x2e];\r
+    hpal[0x80 + 0x0e] = hpal[0x0e];\r
+    hpal[0x80 + 0x1e] = hpal[0x1e];\r
+    hpal[0x80 + 0x2e] = hpal[0x2e];\r
   }\r
 }\r
 \r
@@ -1592,7 +1593,7 @@ static void FinalizeLine8bit(int sh, int line, struct PicoEState *est)
   if (Pico.m.dirtyPal == 1)\r
   {\r
     // a hack for mid-frame palette changes\r
-    if (!(est->rendstatus & PDRAW_SONIC_MODE) || line - dirty_line > 4) {\r
+    if (!(est->rendstatus & PDRAW_SONIC_MODE) | (line - dirty_line > 4)) {\r
       // store a maximum of 2 additional palettes in SonicPal\r
       if (est->SonicPalCount < 2)\r
         est->SonicPalCount ++;\r
@@ -1634,7 +1635,7 @@ static int DrawDisplay(int sh)
   int win=0, edge=0, hvwind=0, lflags;\r
   int maxw, maxcells;\r
 \r
-  est->rendstatus &= ~(PDRAW_SHHI_DONE|PDRAW_PLANE_HI_PRIO|PDRAW_WND_HIGH_PRIO);\r
+  est->rendstatus &= ~(PDRAW_SHHI_DONE|PDRAW_PLANE_HI_PRIO|PDRAW_WND_DIFF_PRIO);\r
 \r
   if (pvid->reg[12]&1) {\r
     maxw = 328; maxcells = 40;\r
@@ -1730,8 +1731,8 @@ static int DrawDisplay(int sh)
 #if 0\r
   {\r
     int *c, a, b;\r
-    for (a = 0, c = HighCacheA; *c; c++, a++);\r
-    for (b = 0, c = HighCacheB; *c; c++, b++);\r
+    for (a = 0, c = HighCacheA; *c; c+=2, a++);\r
+    for (b = 0, c = HighCacheB; *c; c+=2, b++);\r
     printf("%i:%03i: a=%i, b=%i\n", Pico.m.frame_count,\r
            Pico.est.DrawScanline, a, b);\r
   }\r

diff --git a/pico/draw_arm.S b/pico/draw_arm.S
index 7dab089..46ecc77 100644 (file)
--- a/pico/draw_arm.S
+++ b/pico/draw_arm.S
@@ -14,7 +14,7 @@
 .extern DrawStripInterlace\r
 \r
 .equ PDRAW_SPRITES_MOVED, (1<<0)\r
-.equ PDRAW_WND_HIGH_PRIO, (1<<1)\r
+.equ PDRAW_WND_DIFF_PRIO, (1<<1)\r
 .equ PDRAW_PARSE_SPRITES, (1<<2)\r
 .equ PDRAW_DIRTY_SPRITES, (1<<4)\r
 .equ PDRAW_PLANE_HI_PRIO, (1<<6)\r
@@ -426,15 +426,10 @@ DrawLayer:
     add     r1, r1, #8\r
     add     r8, r8, #1\r
 \r
-    tst     r7, #0x8000\r
-    tsteq   r10, #1<<20     @ force?\r
-    bne     .DrawStrip_hiprio\r
-\r
     cmp     r7, r9\r
     beq     .DrawStrip_samecode @ we know stuff about this tile already\r
 \r
     mov     r9, r7          @ remember code\r
-    orr     r10, r10, #1<<21 @ seen non hi-prio tile\r
 \r
     movs    r2, r9, lsl #20 @ if (code&0x1000)\r
     mov     r2, r2, lsl #1\r
@@ -444,14 +439,19 @@ DrawLayer:
 \r
     ldr     r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels\r
 \r
-    bic     r7, r3, #0x7f\r
-    and     r3, r9, #0x6000\r
-    add     r3, r7, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);\r
-\r
 .DrawStrip_samecode:\r
+    tst     r9, #0x8000\r
+    tsteq   r10, #1<<20     @ force?\r
+    bne     .DrawStrip_hiprio\r
+\r
+    orr     r10, r10, #1<<21 @ seen non hi-prio tile\r
     tst     r2, r2\r
     beq     .dsloop              @ tileline blank\r
 \r
+    bic     r7, r3, #0x7f\r
+    and     r3, r9, #0x6000\r
+    add     r3, r7, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);\r
+\r
     cmp     r2, r2, ror #4\r
     beq     .DrawStrip_SingleColor @ tileline singlecolor \r
 \r
@@ -481,13 +481,18 @@ DrawLayer:
     b       .dsloop_subr1\r
 \r
 .DrawStrip_hiprio:\r
-    sub     r0, r1, r11\r
-    orr     r7, r7, r0,  lsl #16\r
+    tst     r10, #(1<<23)   @ sh[23]\r
+    tsteq   r2, r2          @ if (!sh[23] && code==blank) continue\r
+    beq     .dsloop\r
+\r
+@    orr     r10, r10, #1<<22 @ hi_not_empty[22]\r
+    sub     r7, r1, r11\r
+    orr     r7, r9, r7,  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
+    tst     r9, #0x1000\r
+    eorne   r7, r7, #0xe<<25 @ if(code&0x1000) cval^=0xe<<25;\r
+    str     r7, [r6], #4    @ cache hi priority tile code\r
+    str     r2, [r6], #4    @ cache hi priority tile data\r
     b       .dsloop\r
 \r
 .dsloop_exit:\r
@@ -499,8 +504,7 @@ DrawLayer:
     orreq   r2, r2, #PDRAW_PLANE_HI_PRIO @ had a layer with all hi-prio tiles\r
     streq   r2, [r1, #OFS_EST_rendstatus]\r
 \r
-    ldmfd   sp!, {r4-r11,lr}\r
-    bx      lr\r
+    ldmfd   sp!, {r4-r11,pc}\r
 \r
 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@\r
 \r
@@ -593,15 +597,11 @@ DrawLayer:
     add     r1, r1, #8\r
     add     r8, r8, #1\r
 \r
-    tst     r7, #0x8000\r
-    bne     .DrawStrip_vs_hiprio\r
-\r
     orr     r7, r7, r10, lsl #24  @ code | (ty << 24)\r
     cmp     r7, r9\r
     beq     .DrawStrip_vs_samecode @ we know stuff about this tile already\r
 \r
     mov     r9, r7          @ remember code\r
-    orr     r8, r8, #(1<<24)@ seen non hi-prio tile\r
 \r
     movs    r2, r9, lsl #20 @ if (code&0x1000)\r
     mov     r2, r2, lsl #1\r
@@ -611,14 +611,19 @@ DrawLayer:
 \r
     ldr     r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(PicoMem.vram+addr); // Get 8 pixels\r
 \r
-    bic     r7, r3, #0x7f\r
-    and     r3, r9, #0x6000\r
-    add     r3, r7, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);\r
-\r
 .DrawStrip_vs_samecode:\r
+    tst     r9, #0x8000\r
+    tsteq   r10, #(1<<20)      @ force[20]\r
+    bne     .DrawStrip_vs_hiprio\r
+\r
+    orr     r8, r8, #(1<<24)@ seen non hi-prio tile\r
     tst     r2, r2\r
     beq     .dsloop_vs              @ tileline blank\r
 \r
+    bic     r7, r3, #0x7f\r
+    and     r3, r9, #0x6000\r
+    add     r3, r7, r3, lsr #9 @ r3=pal=((code&0x6000)>>9);\r
+\r
     cmp     r2, r2, ror #4\r
     beq     .DrawStrip_vs_SingleColor @ tileline singlecolor \r
 \r
@@ -648,13 +653,18 @@ DrawLayer:
     b       .dsloop_vs_subr1\r
 \r
 .DrawStrip_vs_hiprio:\r
-    sub     r0, r1, r11\r
-    orr     r7, r7, r0,  lsl #16\r
+    tst     r10, #(1<<23)   @ sh[23]\r
+    tsteq   r2, r2          @ if (!sh[23] && code==blank) continue\r
+    beq     .dsloop_vs\r
+\r
+@    orr     r10, r10, #1<<22 @ hi_not_empty[22]\r
+    sub     r7, r1, r11\r
+    orr     r7, r9, r7,  lsl #16\r
     orr     r7, r7, r10, lsl #25 @ (ty<<25)\r
-    tst     r7, #0x1000\r
+    tst     r9, #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
+    str     r7, [r6], #4    @ cache hi priority tile code\r
+    str     r2, [r6], #4    @ cache hi priority tile data\r
     b       .dsloop_vs\r
 \r
 .dsloop_vs_exit:\r
@@ -666,8 +676,7 @@ DrawLayer:
     orreq   r2, r2, #PDRAW_PLANE_HI_PRIO @ had a layer with all hi-prio tiles\r
     streq   r2, [r1, #OFS_EST_rendstatus]\r
 \r
-    ldmfd   sp!, {r4-r11,lr}\r
-    bx      lr\r
+    ldmfd   sp!, {r4-r11,pc}\r
 \r
 \r
 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@\r
@@ -696,8 +705,7 @@ DrawLayer:
     bl      DrawStripInterlace @ struct TileStrip *ts, int plane_sh\r
 \r
     add     sp, sp, #6*4\r
-    ldmfd   sp!, {r4-r11,lr}\r
-    bx      lr\r
+    ldmfd   sp!, {r4-r11,pc}\r
 \r
 .pool\r
 \r
@@ -739,8 +747,7 @@ BackFill:
     stmia   lr!, {r0-r7}\r
     stmia   lr!, {r0-r7}\r
 \r
-    ldmfd   sp!, {r4-r9,lr}\r
-    bx      lr\r
+    ldmfd   sp!, {r4-r9,pc}\r
 \r
 \r
 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@\r
@@ -758,7 +765,6 @@ DrawTilesFromCache:
     ldr     lr, [r3, #OFS_EST_PicoMem_vram]\r
     mov     r9, r3         @ est\r
 \r
-    mvn     r5, #0         @ r5=prevcode=-1\r
     ands    r8, r1, #1\r
     orr     r8, r8, r2, lsl #1\r
     bne     .dtfc_check_rendflags\r
@@ -771,22 +777,11 @@ DrawTilesFromCache:
     bic     r4, r1, #0xfe00\r
     add     r1, r11, r4     @ r1=pdest\r
 \r
-    movs    r7, r6, lsl #16\r
-    cmp     r5, r7, lsr #16\r
-    beq     .dtfc_samecode  @ if (code==prevcode)\r
-\r
-    mov     r5, r7, lsr #16\r
-\r
-    mov     r2, r5, lsl #21\r
-    mov     r2, r2, lsr #17 @ r2=addr=(code&0x7ff)<<4;\r
-    add     r2, r2, r6, lsr #25 @ addr+=ty\r
+    ldr     r2, [r0], #4    @ read pixel data\r
 \r
-    and     r3, r5, #0x6000\r
+    and     r3, r6, #0x6000\r
     mov     r3, r3, lsr #9  @ r3=pal=((code&0x6000)>>9);\r
 \r
-    ldr     r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels\r
-\r
-.dtfc_samecode:\r
     rsbs    r4, r4, r8, lsr #1\r
     bmi     .dtfc_cut_tile\r
 \r
@@ -799,7 +794,7 @@ DrawTilesFromCache:
     cmp     r2, r2, ror #4\r
     beq     .dtfc_SingleColor @ tileline singlecolor \r
 \r
-    tst     r5, #0x0800\r
+    tst     r6, #0x0800\r
     bne     .dtfc_TileFlip\r
 \r
     @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern\r
@@ -831,7 +826,7 @@ DrawTilesFromCache:
     cmp     r2, r2, ror #4\r
     beq     .dtfc_SingleColor @ tileline singlecolor \r
 \r
-    tst     r5, #0x0800\r
+    tst     r6, #0x0800\r
     bne     .dtfc_TileFlipShHP\r
 \r
     @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern\r
@@ -871,7 +866,7 @@ DrawTilesFromCache:
     mov     r12,#0xf<<28\r
     mov     r12,r12,asr r4\r
     mov     r2, r2, ror #16\r
-    tst     r5, #0x0800     @ flipped?\r
+    tst     r6, #0x0800     @ flipped?\r
     mvnne   r12,r12\r
     and     r2, r2, r12\r
     mov     r2, r2, ror #16\r
@@ -880,7 +875,7 @@ DrawTilesFromCache:
     bne     .dtfc_shadow\r
     tst     r2, r2\r
     beq     .dtfc_loop\r
-    tst     r5, #0x0800\r
+    tst     r6, #0x0800\r
     beq     .dtfc_TileNorm\r
     b       .dtfc_TileFlip\r
 \r
@@ -1310,8 +1305,8 @@ DrawWindow:
     @ fetch the first code now\r
     ldrh    r7, [lr, r12]\r
 \r
-    ands    r6, r6, #PDRAW_WND_HIGH_PRIO\r
-    cmpeq   r2, #1                @ prio && !(rendstatus & WND_HIGH_PRIO)?\r
+    ands    r6, r6, #PDRAW_WND_DIFF_PRIO\r
+    cmpeq   r2, #1                @ prio && !(rendstatus & WND_DIFF_PRIO)?\r
     ldmeqfd sp!, {r4-r11,pc}      @ yes, assume that whole window uses same priority\r
 \r
     orr     r6, r6, r2\r
@@ -1339,7 +1334,7 @@ DrawWindow:
 \r
     eor     r5, r6, r7, lsr #15\r
     tst     r5, #1\r
-    orrne   r6, r6, #PDRAW_WND_HIGH_PRIO @ wrong pri\r
+    orrne   r6, r6, #PDRAW_WND_DIFF_PRIO @ wrong pri\r
     bne     .dwloop\r
 \r
     cmp     r7, r9\r
@@ -1405,7 +1400,7 @@ DrawWindow:
     b       .dw_shadow_done\r
 \r
 .dwloop_end:\r
-    and     r2, r6, #PDRAW_WND_HIGH_PRIO\r
+    and     r2, r6, #PDRAW_WND_DIFF_PRIO\r
     ldmfd   sp!, {r4-r11,lr}\r
     ldr     r0, [sp]\r
     ldr     r1, [r0, #OFS_EST_rendstatus]\r
@@ -1463,8 +1458,7 @@ vidConvCpyRGB565: @ void *to, void *from, int pixels
     orr     r8, r8, #0x0800\r
     orr     r8, r8, r8, lsl #16\r
     vidConvCpyRGB565_local\r
-    ldmfd   sp!, {r4-r9,lr}\r
-    bx      lr\r
+    ldmfd   sp!, {r4-r9,pc}\r
 \r
 \r
 @ void PicoDoHighPal555(int sh, int line, struct PicoEState *est)\r
@@ -1532,14 +1526,11 @@ PicoDoHighPal555:
     stmia   r4!, {r1,r6}\r
     bne     .fl_loopcpRGB555_sh\r
 \r
-    @ fixup color 14 in palette 0,1,2 (always normal)\r
+    @ fixup shadowed color 14 in palette 0,1,2 (always normal)\r
     sub     r4, r3, #0x40*2\r
     ldrh    r1, [r4, #0x0e*2] @ 0x0e, 0x1e, 0x2e\r
     ldrh    r5, [r4, #0x1e*2]\r
     ldrh    r6, [r4, #0x2e*2]\r
-    strh    r1, [r3, #0x0e*2] @ 0x4e, 0x5e, 0x6e\r
-    strh    r5, [r3, #0x1e*2]\r
-    strh    r6, [r3, #0x2e*2]\r
     strh    r1, [r3, #0x4e*2] @ 0x8e, 0x9e, 0xae\r
     strh    r5, [r3, #0x5e*2]\r
     strh    r6, [r3, #0x6e*2]\r
@@ -1622,8 +1613,7 @@ FinalizeLine555:
     stmia   r0!, {r4,r5,r8,r12}\r
     bne     .fl_loopRGB555\r
 \r
-    ldmfd   sp!, {r4-r10,lr}\r
-    bx      lr\r
+    ldmfd   sp!, {r4-r10,pc}\r
 \r
 \r
 .fl_32scale_RGB555:\r
@@ -1687,8 +1677,7 @@ FinalizeLine555:
     stmia   r0!, {r4,r5,r6,r8,r10}\r
     bne     .fl_loop32scale_RGB555\r
 \r
-    ldmfd   sp!, {r4-r10,lr}\r
-    bx      lr\r
+    ldmfd   sp!, {r4-r10,pc}\r
 \r
 #ifdef UNALIGNED_DRAWLINEDEST\r
     @ unaligned versions of loops\r
@@ -1733,8 +1722,7 @@ FinalizeLine555:
 \r
     strh    r8, [r0], #2\r
 \r
-    ldmfd   sp!, {r4-r10,lr}\r
-    bx      lr\r
+    ldmfd   sp!, {r4-r10,pc}\r
 \r
 \r
 .fl_32scale_RGB555u:\r
@@ -1799,8 +1787,7 @@ FinalizeLine555:
 \r
     strh    r4, [r0], #2\r
 \r
-    ldmfd   sp!, {r4-r10,lr}\r
-    bx      lr\r
+    ldmfd   sp!, {r4-r10,pc}\r
 \r
 #endif /* UNALIGNED_DRAWLINEDEST */\r
 \r

diff --git a/pico/pico.h b/pico/pico.h
index 6e99045..3c800f6 100644 (file)
--- a/pico/pico.h
+++ b/pico/pico.h
@@ -199,7 +199,7 @@ void vidConvCpyRGB565(void *to, void *from, int pixels);
 void PicoDoHighPal555(int sh, int line, struct PicoEState *est);\r
 // internals, NB must keep in sync with ASM draw functions\r
 #define PDRAW_SPRITES_MOVED (1<<0) // SAT address modified\r
-#define PDRAW_WND_HIGH_PRIO (1<<1) // have window with high prio tiles\r
+#define PDRAW_WND_DIFF_PRIO (1<<1) // not all window tiles use same priority\r
 #define PDRAW_PARSE_SPRITES (1<<2) // SAT needs parsing\r
 #define PDRAW_INTERLACE     (1<<3)\r
 #define PDRAW_DIRTY_SPRITES (1<<4) // SAT modified\r