swizzle address with pre-swapped data instead of endian-swap

bios_data.S

@@ -5,5 +5,5 @@
 .data
 _open_gba_bios_rom:
 open_gba_bios_rom:
-  .incbin "bios/open_gba_bios.bin"
+  .incbin "bios/open_gba_bios.bigendian"
 

common.h

@@ -137,15 +137,21 @@ typedef u32 fixed8_24;
 
 #define eswap8(value)  (value)
 #if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
-  #define eswap16(value) __builtin_bswap16(value)
+//  #define eswap16(value) __builtin_bswap16(value)
-  #define eswap32(value) __builtin_bswap32(value)
+//  #define eswap32(value) __builtin_bswap32(value)
+  #define eswap16(value) (value)
+  #define eswap32(value) (value)
+  #define swizzle_h(ofs) ((ofs) ^ 2)
+  #define swizzle_b(ofs) ((ofs) ^ 3)
 #else
   #define eswap16(value) (value)
   #define eswap32(value) (value)
+  #define swizzle_h(ofs) (ofs)
+  #define swizzle_b(ofs) (ofs)
 #endif
 
-#define  readaddress8(base, offset) eswap8( address8( base, offset))
+#define  readaddress8(base, offset) eswap8(address8( base, swizzle_b(offset)))
-#define readaddress16(base, offset) eswap16(address16(base, offset))
+#define readaddress16(base, offset) eswap16(address16(base, swizzle_h(offset)))
 #define readaddress32(base, offset) eswap32(address32(base, offset))
 
 #define read_ioreg(regnum) (eswap16(io_registers[(regnum)]))

mips/mips_emit.h

@@ -2027,6 +2027,9 @@ u32 execute_store_cpsr_body(u32 _cpsr, u32 address)
 
 #if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
 
+#define emit_swizzle_h(r) mips_emit_xori(r, r, 2)
+#define emit_swizzle_b(r) mips_emit_xori(r, r, 3)
+
     /*
   400230:       00042e00        sll     a1,a0,0x18
   400234:       00041602        srl     v0,a0,0x18
@@ -2076,6 +2079,9 @@ u32 execute_store_cpsr_body(u32 _cpsr, u32 address)
 
 #else
 
+#define emit_swizzle_h()
+#define emit_swizzle_b()
+
 #define emit_eswap32(r)
 #define emit_eswap16(r, sext)
 
@@ -2115,22 +2121,24 @@ static void emit_mem_access_loadop(
   switch (size) {
   case 2:
     mips_emit_lw(reg_rv, reg_rv, (base_addr & 0xffff));
-    emit_eswap32(reg_rv);
     break;
   case 1:
     if (signext) {
       if (alignment) {
         // Unaligned signed 16b load, is just a load byte (due to sign extension)
+        emit_swizzle_b(reg_rv);
         mips_emit_lb(reg_rv, reg_rv, ((base_addr | 1) & 0xffff));
       } else {
+        emit_swizzle_h(reg_rv);
         mips_emit_lh(reg_rv, reg_rv, (base_addr & 0xffff));
       }
     } else {
+      emit_swizzle_h(reg_rv);
       mips_emit_lhu(reg_rv, reg_rv, (base_addr & 0xffff));
     }
-    emit_eswap16(reg_rv, signext);
     break;
   default:
+    emit_swizzle_b(reg_rv);
     if (signext) {
       mips_emit_lb(reg_rv, reg_rv, (base_addr & 0xffff));
     } else {
@@ -2392,8 +2400,10 @@ static void emit_pmemst_stub(
     if (realsize == 2) {
       mips_emit_lw(reg_temp, reg_temp, base_addr);
     } else if (realsize == 1) {
+      emit_swizzle_h(reg_temp);
       mips_emit_lh(reg_temp, reg_temp, base_addr);
     } else {
+      emit_swizzle_b(reg_temp);
       mips_emit_lb(reg_temp, reg_temp, base_addr);
     }
     // If the data is non zero, we just wrote over code
@@ -2403,12 +2413,12 @@ static void emit_pmemst_stub(
 
   // Store the data (delay slot from the SMC branch)
   if (realsize == 2) {
-    emit_eswap32(reg_a1);
     mips_emit_sw(reg_a1, reg_rv, base_addr);
   } else if (realsize == 1) {
-    emit_eswap16(reg_a1, false);
+    emit_swizzle_h(reg_rv);
     mips_emit_sh(reg_a1, reg_rv, base_addr);
   } else {
+    emit_swizzle_b(reg_rv);
     mips_emit_sb(reg_a1, reg_rv, base_addr);
   }
 
@@ -2488,7 +2498,9 @@ static void emit_palette_hdl(
   if (realsize == 2) {
     mips_emit_sw(reg_a1, reg_rv, 0x100);
   } else if (realsize == 1) {
+    emit_swizzle_h(reg_rv);
     mips_emit_sh(reg_a1, reg_rv, 0x100);
+    emit_swizzle_h(reg_rv); // TODO: ???
   }
 
   // Convert and store in mirror memory