From e6eb0a17dcf50e0ff8ff5843362e96020a6de094 Mon Sep 17 00:00:00 2001
From: lif <>
Date: Tue, 2 Jan 2024 03:55:24 -0800
Subject: [PATCH] start fixing swizzle-mode graphics rendering, starting with
 tilemap pointers and OAM

---
 video.cc | 47 +++++++++++++++++++++++++++++++++++++++++------
 1 file changed, 41 insertions(+), 6 deletions(-)

diff --git a/video.cc b/video.cc
index 567b200..d1521b8 100644
--- a/video.cc
+++ b/video.cc
@@ -28,6 +28,26 @@ u32 gba_screen_pitch = GBA_SCREEN_WIDTH;
 #define get_screen_pixels()   gba_screen_pixels
 #define get_screen_pitch()    gba_screen_pitch
 
+#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
+
+typedef struct {
+  u16 attr1, attr0, attr3, attr2;
+} t_oam;
+
+typedef struct {
+  u16 pad0[2];
+  u16 dx;
+  u16 pad1[3];
+  u16 dmx;
+  u16 pad2[3];
+  u16 dy;
+  u16 pad3[3];
+  u16 dmy;
+  u16 pad4;
+} t_affp;
+
+#else // BIG_ENDIAN
+
 typedef struct {
   u16 attr0, attr1, attr2, attr3;
 } t_oam;
@@ -43,6 +63,8 @@ typedef struct {
   u16 dmy;
 } t_affp;
 
+#endif // BIG_ENDIAN
+
 typedef void (* bitmap_render_function)(
   u32 start, u32 end, void *dest_ptr, const u16 *pal);
 typedef void (* tile_render_function)(
@@ -285,6 +307,16 @@ static inline void render_tile_Nbpp(
   }
 }
 
+inline u16 swizzle_u16ptr(u16 *ptr) {
+  u16* swizptr = (u16*)((uintptr_t)(ptr) ^ 2);
+  return *swizptr;
+}
+
+inline u16 swizzle_u16ptr_inc(u16 **pptr) {
+  u16* swizptr = (u16*)((uintptr_t)(*pptr) ^ 2);
+  (*pptr)++;
+  return *swizptr;
+}
 
 template<typename stype, rendtype rdtype, bool isbase, bool is8bpp>
 static void render_scanline_text_fast(u32 layer,
@@ -360,7 +392,10 @@ static void render_scanline_text_fast(u32 layer,
     u32 todraw = MIN(end, partial_hcnt); // [1..7]
     u32 stop = tile_hoff + todraw;   // Usually 8, unless short run.
 
-    u16 tile = eswap16(*map_ptr++);
+//    map_ptr = (u16*)((uintptr_t)(map_ptr) ^ 2);
+    u16 tile = swizzle_u16ptr_inc(&map_ptr); //eswap16(*map_ptr);
+//    map_ptr = (u16*)((uintptr_t)(map_ptr) ^ 2);
+//    map_ptr++;
     if (tile & 0x400)   // Tile horizontal flip
       rend_part_tile_Nbpp<stype, rdtype, is8bpp, isbase, true>(
         bg_comb, px_comb, dest_ptr, tile_hoff, stop, tile, tile_base, vflip_off, paltbl);
@@ -380,7 +415,7 @@ static void render_scanline_text_fast(u32 layer,
   u32 todraw = MIN(end, pixel_run) / 8;
 
   for (u32 i = 0; i < todraw; i++, dest_ptr += 8) {
-    u16 tile = eswap16(*map_ptr++);
+    u16 tile = swizzle_u16ptr_inc(&map_ptr);
     if (tile & 0x400)   // Tile horizontal flip
       render_tile_Nbpp<stype, rdtype, is8bpp, isbase, true>(
         bg_comb, px_comb, dest_ptr, tile, tile_base, vflip_off, paltbl);
@@ -401,7 +436,7 @@ static void render_scanline_text_fast(u32 layer,
 
   todraw = end / 8;
   for (u32 i = 0; i < todraw; i++, dest_ptr += 8) {
-    u16 tile = eswap16(*map_ptr++);
+    u16 tile = swizzle_u16ptr_inc(&map_ptr);//eswap16(*map_ptr++);
     if (tile & 0x400)   // Tile horizontal flip
       render_tile_Nbpp<stype, rdtype, is8bpp, isbase, true>(
         bg_comb, px_comb, dest_ptr, tile, tile_base, vflip_off, paltbl);
@@ -414,7 +449,7 @@ static void render_scanline_text_fast(u32 layer,
 
   // Finalize the tile rendering the left side of it (from 0 up to "end").
   if (end) {
-    u16 tile = eswap16(*map_ptr++);
+    u16 tile = swizzle_u16ptr_inc(&map_ptr);//eswap16(*map_ptr++);
     if (tile & 0x400)   // Tile horizontal flip
       rend_part_tile_Nbpp<stype, rdtype, is8bpp, isbase, true>(
         bg_comb, px_comb, dest_ptr, 0, end, tile, tile_base, vflip_off, paltbl);
@@ -488,7 +523,7 @@ static void render_scanline_text_mosaic(u32 layer,
   // Iterate pixel by pixel, loading data every N pixels to honor mosaic effect
   u8 pval = 0;
   for (u32 i = 0; start < end; start++, i++, dest_ptr++) {
-    u16 tile = eswap16(*map_ptr);
+    u16 tile = swizzle_u16ptr(map_ptr);//eswap16(*map_ptr);
 
     if (!(i % mosh)) {
       const u8 *tile_ptr = &tile_base[(tile & 0x3FF) * (is8bpp ? 64 : 32)];
@@ -1431,7 +1466,7 @@ void render_scanline_objs(
   for (objn = objcnt-1; objn >= 0; objn--) {
     // Objects in the list are pre-filtered and sorted in the appropriate order
     u32 objoff = objlist[objn];
-    const t_oam *oamentry = &((t_oam*)oam_ram)[objoff];
+    const t_oam *oamentry = &((t_oam*)oam_ram)[objoff ^ (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)];
 
     u16 obj_attr0 = eswap16(oamentry->attr0);
     u16 obj_attr1 = eswap16(oamentry->attr1);