Final cleanups for Fast mode

src/Tiles.s

@@ -352,14 +352,12 @@ _SetTileProcs
 FastProcs
 FastOverZA   dw   _TBConstTile0,GenericOverZero,_OneSpriteFastOver0
 FastOverZV   dw   _TBConstTile0,GenericOverZero,_OneSpriteFastOver0
-FastOverNA   dw   _TBCopyDataAFast,GenericOverZero,_OneSpriteFastOverA
+FastOverNA   dw   _TBCopyDataAFast,GenericOverAFast,_OneSpriteFastOverA
-FastOverNV   dw   _TBCopyDataVFast,GenericOverZero,_OneSpriteFastOverV
+FastOverNV   dw   _TBCopyDataVFast,GenericOverVFast,_OneSpriteFastOverV
-;FastOverNA   dw   _TBCopyDataAFast,GenericOverAFast,_OneSpriteFastOverA
-;FastOverNV   dw   _TBCopyDataVFast,GenericOverVFast,_OneSpriteFastOverV
 FastUnderZA  dw   _TBConstTile0,GenericUnderZero,GenericUnderZero
 FastUnderZV  dw   _TBConstTile0,GenericUnderZero,GenericUnderZero
-FastUnderNA  dw   _TBConstTile0,GenericOverZero,_OneSpriteFastUnderA
+FastUnderNA  dw   _TBCopyDataAFast,GenericUnderAFast,_OneSpriteFastUnderA
-FastUnderNV  dw   _TBConstTile0,GenericOverZero,_OneSpriteFastOver0
+FastUnderNV  dw   _TBCopyDataVFast,GenericUnderVFast,_OneSpriteFastUnderV
 ;FastUnderNA  dw   _TBCopyDataFast,GenericUnderAFast,_OneSpriteFastUnderA
 ;FastUnderNV  dw   _TBCopyDataVFast,GenericUnderVFast,_OneSpriteFastUnderV
 

src/render/Fast.s

@@ -90,12 +90,12 @@ GenericUnderAFast
             lup   8
             lda   tmp_sprite_data+{]line*4}
             andl  tiledata+{]line*4}+32,x
-            oral  tiledata+{]line*4}+32,x
+            oral  tiledata+{]line*4},x
             sta:  $0004+{]line*$1000},y
 
             lda   tmp_sprite_data+{]line*4}+2
             andl  tiledata+{]line*4}+32+2,x
-            oral  tiledata+{]line*4}+32+2,x
+            oral  tiledata+{]line*4}+2,x
             sta:  $0001+{]line*$1000},y
 ]line       equ   ]line+1
             --^
@@ -116,12 +116,12 @@ GenericUnderVFast
             lup   8
             lda   tmp_sprite_data+{]line*4}
             andl  tiledata+{]src*4}+32,x
-            oral  tiledata+{]src*4}+32,x
+            oral  tiledata+{]src*4},x
             sta:  $0004+{]line*$1000},y
 
             lda   tmp_sprite_data+{]line*4}+2
             andl  tiledata+{]src*4}+32+2,x
-            oral  tiledata+{]src*4}+32+2,x
+            oral  tiledata+{]src*4}+2,x
             sta:  $0001+{]line*$1000},y
 ]src        equ   ]src-1
 ]dest       equ   ]dest+1

src/render/Render.s

@@ -22,17 +22,17 @@ _RenderTile
 ; flatten the sprite data into the direct page workspace and then pass control
 ; to the configurable routine which is set in SetTile and knows what to do
 ; based on the tile properties (over/under, engine mode, etc.)
+;
+; NOTE: Could pull the CopyXXXSprites function inline and same the 3 cycles for the JMP,
+;       - or - put the TYX into the macro and jump directly from there.
 TwoSprites  tyx
-            jsr   CopyTwoSpritesDataAndMaskToDP
+            jmp   CopyTwoSpritesDataAndMaskToDP
-            jmp   (K_TS_SPRITE_TILE_DISP,x)
 
 ThreeSprites tyx
-            jsr   CopyThreeSpritesDataAndMaskToDP
+            jmp   CopyThreeSpritesDataAndMaskToDP
-            jmp   (K_TS_SPRITE_TILE_DISP,x)
 
 FourSprites tyx
-            jsr   CopyFourSpritesDataAndMaskToDP
+            jmp   CopyFourSpritesDataAndMaskToDP
-            jmp   (K_TS_SPRITE_TILE_DISP,x)
 
 ; Helper functions (and macros)
 

src/render/Sprite1.s

@@ -27,11 +27,8 @@ _OneSpriteFastOver0
             plb                                    ; Restore the TileStore bank
             rts
 
-; Next implementation; drawing a sprite onto a regular tile.  In this case we need to make
+; Next implementation; drawing a sprite onto a regular tile. The 1-sprite dispatch preerves the
-; use of the K_TS_COPY_TILE_DATA function because that takes care of copying the correct
+; X-register, so it already points to the TileStore
-; tile data into the direct page buffer.
-
-; The 1-sprite dispatch prserves the X-register, so it already points to the TileStore
 
 _OneSpriteFastOverV
             jsr   FastCopyTileDataV

src/render/Sprite2.s

@@ -80,9 +80,13 @@ CopyTwoSpritesDataToDP
 ]line       equ   ]line+1
             --^
             rts
+
 CopyFourSpritesDataAndMaskToDP
 CopyThreeSpritesDataAndMaskToDP
 CopyTwoSpritesDataAndMaskToDP
+            pei   DP2_SPRITEDATA_AND_TILESTORE_BANKS
+            plb
+
 ]line       equ   0
             lup   8
             ldy   #{]line*SPRITE_PLANE_SPAN}
@@ -98,5 +102,7 @@ CopyTwoSpritesDataAndMaskToDP
             sta   tmp_sprite_mask+{]line*4}+2
 ]line       equ   ]line+1
             --^
+
+            plb
             jmp   (K_TS_SPRITE_TILE_DISP,x)