Save another ~45 sec runtime, version 0.1.1

1 files changed, 33 insertions, 22 deletions

diff --git a/dla.s b/dla.s
index 4a30921..48a6031 100644
--- a/dla.s
+++ b/dla.s
@@ -33,11 +33,7 @@ pixptr:       .res 2 ; used by plotsetup and friends
 pixmask:      .res 1 ; ditto.
 cursor_x:     .res 1 ; cursor x/y are args to plot/unplot/locate
 cursor_y:     .res 1
-
-min_x:        .res 1 ; limits: if the particle gets outside this box,
-max_x:        .res 1 ; delete it and spawn a new one.
-min_y:        .res 1
-max_y:        .res 1
+pixptr2:      .res 2 ; used by drunkwalk
 
 circlesize:   .res 1 ; 0 to 3
 
@@ -441,7 +437,7 @@ isloop:
 ;;; - set pixmask to the mask for cursor_x.
 ;;; - set Y reg to the byte offset for cursor_x.
 ;;; - returns with cursor_x in X reg, pixmask in A reg too.
-;;; Called by plot, unplot, and drunkwalk (a lot!)
+;;; Called by plot, unplot, and locate.
 plotsetup:
  ldx cursor_y
  lda lineaddrs_l,x
@@ -508,7 +504,13 @@ oob:
 ;;; possible (we're not there yet).
 drunkwalk:
  ; X holds the X coord the whole time, only needs to be loaded on entry.
- ldx part_x ;3
+ ; preload pixptr, too.
+ ldx part_x
+ ldy part_y ;3
+ lda lineaddrs_l,y    ; 5
+ sta pixptr           ; 3
+ lda lineaddrs_h,y    ; 5
+ sta pixptr+1         ; 3
  ; using bit/bmi/bvc saves 5.25 cycles on average, compared to
  ; immediate cmp and bne.
  ; 4 code paths: up=15, down=18, left=19, right=17, avg=17.25.
@@ -516,7 +518,7 @@ drunkwalk:
  ; are "branch always".
  ; all the "cmp #0" here get their operands modified by set_limits.
 dwloop:
- ldy part_y
+ ldy part_y ; 3
  bit RANDOM ;4 ; use top 2 bits (probably more random, definitely faster)
  bmi lr     ;2/3
  bvc down   ;2/3
@@ -537,15 +539,25 @@ lr:
 selfmod_xmin = * + 1
  cpx #0               ; 2
  beq oob              ; 2
- bne checkneigh ;3
+ ldy xoffsets-1,x ; moved left, check left X neighbor only.
+ lda xmasks-1,x   ; right X neighbor definitely empty, because
+ and (pixptr),y   ; we just moved out of that cell.
+ bne stick
+ beq check_y  ; 3 ; still have to check Y (up/down) neighbors.
 right:
  inx        ;3 ; N=0 V=0 right
 selfmod_xmax = * + 1
  cpx #0               ; 2
  beq oob              ; 2
+ ldy xoffsets+1,x ; as above, moved right, check right neighbor only.
+ lda xmasks+1,x
+ and (pixptr),y
+ bne stick
+ beq check_y          ; 3
 
 checkneigh:
  ; check neighbors. used to be a subroutine, inlined it.
+ ; we only get here when the pixel has moved up or down (not left/right).
  ; also inlined plotsetup here.
  sty part_y           ; 3
  lda lineaddrs_l,y    ; 5
@@ -574,34 +586,33 @@ slow_x:
  lda xmasks+1,x
  and (pixptr),y
  bne stick
+
 check_y:
+ ; this happens no matter what direction the pixel moved.
  ; (0,-1)
- ; subtract 32 (one line) from the pointer. slightly faster
+ ; subtract 32 (one line) from the pointer. one cycle faster
  ; than reloading from lineaddrs_l/h table.
  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:
+ sta pixptr2  ; 3
+ lda pixptr+1 ; 3
+ sbc #0       ; 2
+ sta pixptr2+1 ; 3
  ;ldx part_x ; X already has this from before
  ldy xoffsets,x
  lda xmasks,x
  sta pixmask
- and (pixptr),y
+ and (pixptr2),y
  bne stick
  ; (0,1)
  tya
  ora #$40 ; add 64, AKA 2 screen lines
  tay
- lda (pixptr),y
+ lda (pixptr2),y
  and pixmask
- beq dwloop ; ...or fall through to stick.
- ; note that if we add much more code to drunkwalk, the beq will
- ; have to become a jmp, which takes 3 extra cycles. so any code
- ; added above had better save more than 3 cycles!
+ bne stick
+ jmp dwloop ; too far for a branch
 
 stick: ; we always get here with Z flag clear
  stx part_x ; only update part_x at exit.
@@ -761,7 +772,7 @@ ci_done:
 ; banner and saveprompt must start with a clear-screen code.
 banner:
  .byte $7d, "Diffusion Limited Aggregate",$9b
- .byte "Urchlay's ASM version 0.1.0",$9b,$9b
+ .byte "Urchlay's ASM version 0.1.1",$9b,$9b
  .byte "Particle count range: 1 to 65535",$9b
  .byte "How many particles [",$0