Micro-optimization: Self-modifying code!

1 files changed, 25 insertions, 9 deletions

diff --git a/dla.s b/dla.s
index 4b2be32..4b861f8 100644
--- a/dla.s
+++ b/dla.s
@@ -380,13 +380,13 @@ jsaveimage:
 set_limits:
  ldx circlesize
  lda xmin,x
- sta min_x
+ sta selfmod_xmin
  lda ymin,x
- sta min_y
+ sta selfmod_ymin
  lda xmax,x
- sta max_x
+ sta selfmod_xmax
  lda ymax,x
- sta max_y
+ sta selfmod_ymax
  rts
 
 ;;; Subroutine: initscreen
@@ -526,16 +526,21 @@ right:
  inc part_x ;3
 
 checkbounds:
+ ; all the "cmp #0" here get their operands modified by set_limits.
  lda part_x
- cmp min_x
+selfmod_xmin = * + 1
+ cmp #0
  beq oob
- cmp max_x
+selfmod_xmax = * + 1
+ cmp #0
  beq oob
  ;sta cursor_x
  lda part_y
- cmp min_y
+selfmod_ymin = * + 1
+ cmp #0
  beq oob
- cmp max_y
+selfmod_ymax = * + 1
+ cmp #0
  beq oob
  ;sta cursor_y
  ; checkbounds is 30 cycles when the pixel is in bounds.
@@ -573,12 +578,23 @@ slow_x:
  bne stick
 check_y:
  ; (0,-1)
+.if 0
  ldx part_y        ;3
  lda lineaddrs_l-1,x ;5
  sta pixptr          ;3
  lda lineaddrs_h-1,x ;5
  sta pixptr+1        ;3
                      ;=19
+.else
+ lda pixptr   ; 3
+ sec          ; 2
+ sbc #$20     ; 2
+ sta pixptr   ; 3
+ bcs pp1ok    ; 3|2
+ dec pixptr+1 ; 0|5
+              ; =13|17 (avg closer to 13)
+pp1ok:
+.endif
  ldx part_x
  ldy xoffsets,x
  lda xmasks,x
@@ -594,7 +610,7 @@ check_y:
  bne stick
  jmp drunkwalk ; too far for a branch
 
-stick:
+stick: ; we always get here with Z flag clear
  rts
 
 ;;; Subroutine: drawseed