From 3c41e97b0fe717d07e5f925ab0a518d7f5a5bca4 Mon Sep 17 00:00:00 2001
From: Lucas Scharenbroich <lscharen@prowestgis.com>
Date: Wed, 8 Jun 2022 17:34:23 -0500
Subject: [PATCH] Move dispatch table into K bank and save 9 cycles per tile in
 dispatch

---
 src/Core.s               |  1 -
 src/Tiles.s              | 11 ++++++++---
 src/tiles/FastRenderer.s | 17 ++++++-----------
 3 files changed, 14 insertions(+), 15 deletions(-)

diff --git a/src/Core.s b/src/Core.s
index 619f4ac..f102699 100644
--- a/src/Core.s
+++ b/src/Core.s
@@ -46,7 +46,6 @@ ToolStartUp
                   rts
 
 MasterId          ds        2
-;UserId            ds        2
 
 ; Fatal error handler invoked by the _Err macro
 PgmDeath          tax
diff --git a/src/Tiles.s b/src/Tiles.s
index d423e49..8a24c61 100644
--- a/src/Tiles.s
+++ b/src/Tiles.s
@@ -111,7 +111,7 @@ InitTiles
                  bra  :out
 :fast
                  ldal FastTileProcs
-                 sta  TileStore+TS_BASE_TILE_DISP,x
+                 stal K_TS_BASE_TILE_DISP,x
 :out
 
 ;                 lda  DirtyTileProcs                    ; Fill in with the first dispatch address
@@ -213,7 +213,9 @@ _SetTile
 
 :fast
                  ldal FastTileProcs,x
-                 sta  TileStore+TS_BASE_TILE_DISP,y
+                 tyx
+                 stal K_TS_BASE_TILE_DISP,x
+
 :out
                  jmp  _PushDirtyTileY               ; on the next call to _ApplyTiles
 
@@ -412,4 +414,7 @@ b_12_3     stpbit 12;3;]4
 b_13_3     stpbit 13;3;]4
 b_14_3     stpbit 14;3;]4
 b_15_3     endbit 15;3;]4
-           <<<
\ No newline at end of file
+           <<<
+
+; Store some tables in the K bank that will be used exclusively for jmp (abs,x) dispatch
+K_TS_BASE_TILE_DISP ds TILE_STORE_SIZE
\ No newline at end of file
diff --git a/src/tiles/FastRenderer.s b/src/tiles/FastRenderer.s
index 00b46af..4363625 100644
--- a/src/tiles/FastRenderer.s
+++ b/src/tiles/FastRenderer.s
@@ -13,25 +13,20 @@ NoSpriteFast
             lda   TileStore+TS_CODE_ADDR_HIGH,x    ; load the bank of the target code field line
             pha                                    ; and put on the stack for later. Has TileStore bank in high byte.
             ldy   TileStore+TS_CODE_ADDR_LOW,x     ; load the address of the code field
-            lda   TileStore+TS_BASE_TILE_DISP,x    ; go to the tile copy routine (just basics)
-            stal  nsf_patch+1
+;            lda   TileStore+TS_BASE_TILE_DISP,x    ; go to the tile copy routine (just basics)
+;            stal  nsf_patch+1
             lda   TileStore+TS_TILE_ADDR,x         ; load the address of this tile's data (pre-calculated)
             plb                                    ; set the code field bank
-nsf_patch   jmp   $0000
+            jmp   (K_TS_BASE_TILE_DISP,x)
+
+;nsf_patch   jmp   $0000
 
 ; The TS_BASE_TILE_DISP routines will come from this table when ENGINE_MODE_TWO_LAYER and
 ; ENGINE_MODE_DYN_TILES are both off.
 FastTileProcs dw   _TBCopyDataFast,_TBCopyDataFast,_TBCopyDataFast,_TBCopyDataFast
+
 ; dw   _TBCopyDataFast,_TBCopyDataFast,_TBCopyDataVFast,_TBCopyDataVFast
 
-; NOTE: Inlining the dispatch would eliminate a JSR,RTS,LDX, and JMP (abs,x) because the exit code
-;       could jump directly to the target address.  Net savings of 20 cycles per tile.  For a 16x16
-;       sprite with a 3x3 block coverage this is 180 cycles per frame per block...  This would also 
-;       preserve a register
-;
-;       For comparison, a fast one sprite copy takes 22 cycles per word, so this would save
-;       about 1/2 block of render time per tile.
-;
 ; Need to determine if the sprite or tile data is on top, as that will decide whether the
 ; sprite or tile data is copied into the temporary buffer first.  Also, if TWO_LAYER is set
 ; then the mask information must be copied as well....This is the last decision point.