Fix cycle-counting comments.

1 files changed, 11 insertions, 16 deletions

diff --git a/dla.s b/dla.s
index 4b861f8..229a9ef 100644
--- a/dla.s
+++ b/dla.s
@@ -507,23 +507,25 @@ oob:
 ;;; This is the innermost loop, so it should be as optimized as
 ;;; possible (we're not there yet).
 drunkwalk:
- ; using bit/bmi/bvc saves 6.25 cycles on average, compared to
+ ; using bit/bmi/bvc saves 4.25 cycles on average, compared to
  ; immediate cmp and bne.
- ; 4 code paths: up=14, down=15, left=15, right=13, avg=14.25
+ ; 4 code paths: up=16, down=17, left=17, right=15, avg=16.25.
+ ; note that part_x and part_y are *never* zero; all the bne's here
+ ; are "branch always".
  bit RANDOM ;4 ; use top 2 bits (probably more random, definitely faster)
  bmi lr     ;2/3
  bvc down   ;2/3
- dec part_y ;3 ; N=1 V=1 up
+ dec part_y ;5 ; N=1 V=1 up
  bne checkbounds ;3
 down:
- inc part_y ;3 ; N=1 V=0 down
+ inc part_y ;5 ; N=1 V=0 down
  bne checkbounds ;3
 lr:
  bvc right ;2/3
- dec part_x ;3 ; N=0 V=1 left
+ dec part_x ;5 ; N=0 V=1 left
  bne checkbounds ;3
 right:
- inc part_x ;3
+ inc part_x ;5
 
 checkbounds:
  ; all the "cmp #0" here get their operands modified by set_limits.
@@ -578,14 +580,8 @@ 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
+ ; subtract 32 (one line) from the pointer. slightly faster
+ ; than reloading from lineaddrs_l/h table.
  lda pixptr   ; 3
  sec          ; 2
  sbc #$20     ; 2
@@ -594,8 +590,7 @@ check_y:
  dec pixptr+1 ; 0|5
               ; =13|17 (avg closer to 13)
 pp1ok:
-.endif
- ldx part_x
+ ;ldx part_x ; X already has this from before
  ldy xoffsets,x
  lda xmasks,x
  sta pixmask