From 915a82587d16334cd65a187659315b077364e3c8 Mon Sep 17 00:00:00 2001
From: Peter Ferrie <peter.ferrie@gmail.com>
Date: Tue, 22 Oct 2019 22:15:33 -0700
Subject: [PATCH] a bit faster

---
 src/fx/fx.hgr.precomputed.1bit.a | 46 ++++++-----------
 src/fx/fx.hgr.ripple.a           |  3 ++
 src/fx/macros.a                  | 84 +++++++++++++++++++++++++++++++-
 3 files changed, 101 insertions(+), 32 deletions(-)

diff --git a/src/fx/fx.hgr.precomputed.1bit.a b/src/fx/fx.hgr.precomputed.1bit.a
index 9bc2a7c6d..94768ce43 100644
--- a/src/fx/fx.hgr.precomputed.1bit.a
+++ b/src/fx/fx.hgr.precomputed.1bit.a
@@ -16,9 +16,10 @@ sourcemasks    = $0201    ; same as copymasks
 mirror_copymasks = $0202
 mirror_sourcemasks = $0203
 hgrlo          = $0300    ; $C0 bytes
-mirror_cols    = $BD58    ; $28 bytes
-mirror_rows    = $BD80    ; $C0 bytes
+hgrlomirror    = $BD40    ; $C0 bytes
+mirror_cols    = $BE00    ; $28 bytes
 hgr1hi         = $BE40    ; $C0 bytes
+hgr1himirror   = $BF40    ; $C0 bytes
 
          !source "src/fx/macros.a"
 
@@ -35,6 +36,7 @@ hgr1hi         = $BE40    ; $C0 bytes
 !macro FX_PRECOMPUTED_1BIT .coords {
          ; build lookup tables for base address of HGR rows
          +BUILD_HGR_LOOKUP_TABLES hgrlo, hgr1hi
+         +BUILD_HGR_MIRROR_LOOKUP_TABLES hgrlomirror, hgr1himirror
 
          ; build lookup table to get $27-y for y in $00..$27
          ldx   #$27
@@ -45,16 +47,6 @@ hgr1hi         = $BE40    ; $C0 bytes
          dex
          bpl   -
 
-         ; build lookup table to get $BF-x for x in $00..$BF
-         ldx   #$BF
-         ldy   #$00
--        tya
-         sta   mirror_rows,x
-         dex
-         iny
-         cpy   #$C0
-         bne   -
-
          ; build sparse lookup tables for bitmasks
          lda   #%10000001
          sta   copymasks
@@ -111,9 +103,6 @@ hgr1hi         = $BE40    ; $C0 bytes
          lda   #>.coords
          sta   input+1
 
-         jmp   InputLoop
-Exit1    rts
-
 InputLoop
          ldy   #0
          lda   (input),y             ; first value: HGR row (only 0..95 will be in input array)
@@ -124,7 +113,8 @@ InputLoop
          sta   src1
          lda   hgr1hi,x
          sta   dest1+1
-         eor   #$60
+Exit1 = * + 1
+         eor   #$60                  ; ahem. doubles as RTS
          sta   src1+1
          lda   hgrlo+1,x
          sta   dest2
@@ -134,19 +124,17 @@ InputLoop
          eor   #$60
          sta   src2+1
 
-         lda   mirror_rows,x
-         tax
-         lda   hgrlo,x
+         lda   hgrlomirror,x
          sta   dest3
          sta   src3
-         lda   hgr1hi,x
+         lda   hgr1himirror,x
          sta   dest3+1
          eor   #$60
          sta   src3+1
-         lda   hgrlo-1,x
+         lda   hgrlomirror+1,x
          sta   dest4
          sta   src4
-         lda   hgr1hi-1,x
+         lda   hgr1himirror+1,x
          sta   dest4+1
          eor   #$60
          sta   src4+1
@@ -182,15 +170,13 @@ InputLoop
          +COPY_BIT src3, dest3, sourcemasks, copymasks
          +COPY_BIT src4, dest4, sourcemasks, copymasks
 
-         bit   $c000
-         bmi   Exit2
-
          inc   input
-         bne   +
+         inc   input
+         beq   +
+         jmp   InputLoop
++        bit   $c000
+         bmi   Exit2
          inc   input+1
-+        inc   input
-         +LBNE InputLoop
-         inc   input+1
-         +LBNE InputLoop
+         jmp   InputLoop
 Exit2
 }
diff --git a/src/fx/fx.hgr.ripple.a b/src/fx/fx.hgr.ripple.a
index 3c19e4007..186f47139 100644
--- a/src/fx/fx.hgr.ripple.a
+++ b/src/fx/fx.hgr.ripple.a
@@ -11,6 +11,9 @@
 
          rts
 
+!if * and $fff {
+  !fill $1000 - (* and $fff) ;align 4096
+}
 Coordinates
          !source "src/fx/fx.hgr.ripple.data.a"
          !byte $80
diff --git a/src/fx/macros.a b/src/fx/macros.a
index b32662774..277e35b5f 100644
--- a/src/fx/macros.a
+++ b/src/fx/macros.a
@@ -22,14 +22,94 @@
          rol
          asl   .hgrlo,x
          rol
-         clc
-         adc   #$20
+         ora   #$20
          sta   .hgr1hi,x
          inx
          cpx   #$C0
          bne   -
 }
 
+; .hgrlo, .hgr1hi, .hgr2hi will each be filled with $C0 bytes
+!macro BUILD_HGR12_LOOKUP_TABLES .hgrlo, .hgr1hi, .hgr2hi {
+         ldx   #0
+-        txa
+         and   #$F8
+         bpl   +
+         ora   #5
++        asl
+         bpl   +
+         ora   #5
++        asl
+         asl
+         sta   .hgrlo,x
+         txa
+         and   #7
+         rol
+         asl   .hgrlo,x
+         rol
+         ora   #$20
+         sta   .hgr1hi,x
+         eor   #$60
+         sta   .hgr2hi,x
+         inx
+         cpx   #$C0
+         bne   -
+}
+
+; .hgrlomirror, .hgr1himirror will each be filled with $C0 bytes
+!macro BUILD_HGR_MIRROR_LOOKUP_TABLES .hgrlomirror, .hgr1himirror {
+         ldx   #$C0
+         ldy   #0
+-        tya
+         and   #$F8
+         bpl   +
+         ora   #5
++        asl
+         bpl   +
+         ora   #5
++        asl
+         asl
+         sta   .hgrlomirror-1,x
+         tya
+         and   #7
+         rol
+         asl   .hgrlomirror-1,x
+         rol
+         ora   #$20
+         sta   .hgr1himirror-1,x
+         iny
+         dex
+         bne   -
+}
+
+; .hgrlomirror, .hgr1himirror, .hgr2himirror will each be filled with $C0 bytes
+!macro BUILD_HGR_MIRROR_LOOKUP_TABLES .hgrlomirror, .hgr1himirror, .hgr2mirror {
+         ldx   #$C0
+         ldy   #0
+-        tya
+         and   #$F8
+         bpl   +
+         ora   #5
++        asl
+         bpl   +
+         ora   #5
++        asl
+         asl
+         sta   .hgrlomirror-1,x
+         tya
+         and   #7
+         rol
+         asl   .hgrlomirror-1,x
+         rol
+         ora   #$20
+         sta   .hgr1himirror-1,x
+         eor   #$60
+         sta   .hgr2himirror-1,x
+         iny
+         dex
+         bne   -
+}
+
 !macro HGR_CALC {
          asl
          tax