Small optimization to save 1 cycle per line

src/blitter/PEISlammer.s

@@ -2,6 +2,14 @@
 ; this implementation does is that it does break up the slam into chunks of scan lines to allow
 ; time for interrupts to be serviced in a timely manner.
 ;
+; This is a fairly basic slam in that it does not try to align the direct page.  To enhance the
+; slammer, note that page-aligned addresses repeat every 8 scan lines and some lines would need
+; to be split into two slams to keep the direct page aligned.
+;
+; At best, this saves 1 cycles per word, or 80 cycles for a full screen -- which is only about
+; 12 additional instructions, so this is an optimization that is unlikely to lead to a net
+; improvement.
+;
 ; A = base address of top-left edge of the screen
 ; Y = number of scanlines to blit
 ; X = width of the screen in bytes
@@ -23,11 +31,14 @@ PEISlam
                sta   :inner+1
 
                clc                  ; clear before the loop -- nothing in the loop affect the carry bit
-               bra   :outer         ; hop into the entry point.  The loop control logic is next because 
-                                    ; the size of the PEI instruction is too large to use short branches
-                                    ; in the code after pei_end
+               brl   :outer         ; hop into the entry point.
 
-:control
+]dp            equ   158
+               lup   80             ; A full width screen is 160 bytes / 80 words
+               pei   ]dp
+]dp            equ   ]dp-2
+               --^
+:pei_end
                tdc                  ; Move to the next line
                adc   #160
                tcd
@@ -37,9 +48,20 @@ PEISlam
                dey                  ; decrement the total counter, if zero then we're done
                beq   :exit
 
-               dex                  ; decrement the inner counter
-               bne   :inner         ; if not zero, no break; go to the next line
+               dex                  ; decrement the inner counter.  Both counters are set
+               beq   :restore       ; up so that they fall-through by default to save a cycle
+                                    ; per loop iteration.
 
+:inner         jmp   $0000          ; 25 cycles of overhead per line. A full width slam executes all
+                                    ; 80 of the PEI instructions which we expect to take 7 cycles
+                                    ; since the direct page is not aligned.  So total overhead is
+                                    ; 25 / (25 + 7 * 80) = 4.27% of execution
+                                    ;
+                                    ; Without the interrupt breaks, we could remove the dex/beq test
+                                    ; and save 4 cycles per loop which takes the overhead down to
+                                    ; only 3.6%
+
+:restore
                tsx                  ; save the current stack
                _R0W0                ; restore the execution environment and
                lda   :stk_save      ; give a few cycles to catch some interrupts
@@ -51,7 +73,7 @@ PEISlam
                txs                  ; set the stack address to the right edge
                ldx   #8             ; Enable interrupts at least once every 8 lines
                _R1W1
-:inner         jmp   $0000
+               bra   :inner
 
 :exit
                _R0W0
@@ -62,14 +84,6 @@ PEISlam
                pld
                rts
 
-]dp            equ   158
-               lup   80             ; A full width screen is 160 bytes / 80 words
-               pei   ]dp
-]dp            equ   ]dp-2
-               --^
-:pei_end
-               jmp   :control
-
 :stk_save      ds    2
 :screen_width  ds    2
 
@@ -108,6 +122,19 @@ PEISlam
 
 
 
+
+
+
+
+
+
+
+
+
+
+
+
+