tfv: more optimization, move SCREEN_X to register X

gr-sim/tfv_flying_6502.c

@@ -109,16 +109,14 @@ struct cycle_counts {
 	int put_sprite;
 } cycles;
 
-static int last_color=0,last_xx=0,last_yy=0;
-
 static int lookup_map(int xx, int yy) {
 
 	int color,offset;
 
-	last_xx=xx;
+	ram[LAST_SPACEX_I]=xx;
 	xx=xx&MASK_X;
 
-	last_yy=yy;
+	ram[LAST_SPACEY_I]=yy;
 	yy=yy&MASK_Y;
 
 
@@ -132,7 +130,7 @@ static int lookup_map(int xx, int yy) {
 
 	offset=yy<<3;
 	offset+=xx;
-				cycles.lookup_map+=37;
+				cycles.lookup_map+=39;
 
 	if ((yy>7) || (xx>7)) {
 			cycles.lookup_map+=14;
@@ -156,8 +154,8 @@ static int lookup_map(int xx, int yy) {
 
 
 update_cache:
-				cycles.lookup_map+=9;
+				cycles.lookup_map+=10;
-	last_color=color;
+	ram[LAST_MAP_COLOR]=color;
 	return color;
 }
 
@@ -789,29 +787,18 @@ static void draw_background_mode7(void) {
 		fixed_add(ram[SPACEY_I],ram[SPACEY_F],
 			ram[TEMP_I],ram[TEMP_F],
 			&ram[SPACEY_I],&ram[SPACEY_F]);
-							cycles.mode7+=25;
+							cycles.mode7+=22;
 		if (!displayed) {
 			printf("SPACEY! %x:%x\n",ram[SPACEY_I],ram[SPACEY_F]);
 		}
 
 		// go through all points in this screen line
-		for (ram[SCREEN_X] = 0;
+		ram[SCREEN_X] = 0;
-			ram[SCREEN_X] < LOWRES_W;
-			ram[SCREEN_X]++) {
 
+		while(1) {
 			// get a pixel from the tile and put it on the screen
 
-			/* cache last value */
+nomatch:
-						cycles.mode7+=9;
-			if (ram[SPACEY_I]==last_yy) {
-						cycles.mode7+=8;
-				if (ram[SPACEX_I]==last_xx) {
-						cycles.mode7+=6;
-					map_color=last_color;
-					goto match;
-				}
-			}
-
 			map_color=lookup_map(ram[SPACEX_I],ram[SPACEY_I]);
 						cycles.mode7+=6;
 match:
@@ -839,7 +826,24 @@ match:
 			fixed_add(ram[SPACEY_I],ram[SPACEY_F],
 				ram[DY_I],ram[DY_F],
 				&ram[SPACEY_I],&ram[SPACEY_F]);
-							cycles.mode7+=49;
+
+			ram[SCREEN_X]++;
+			if (ram[SCREEN_X] >= LOWRES_W) break;
+							cycles.mode7+=43;
+
+
+			/* cache last value */
+						cycles.mode7+=8;
+			if (ram[SPACEY_I]==ram[LAST_SPACEY_I]) {
+						cycles.mode7+=7;
+				if (ram[SPACEX_I]==ram[LAST_SPACEX_I]) {
+						cycles.mode7+=4;
+					map_color=ram[LAST_MAP_COLOR];
+					goto match;
+				}
+			}
+			goto nomatch;
+
 		}
 
 		ram[SCREEN_Y]+=1;

tfv/OPTIMIZATION

@@ -146,5 +146,17 @@ Add some self-modifying code to inner loop:
 	Total =                 205,761
 	Frame Rate = 4.86 fps
 
+More self-modifying code, also move SCREEN_X to X register
+	Cycles: flying=             187
+	Cycles: getkey=              46
+	Cycles: page_flip=           26
+	Cycles: multiply=        49,613
+	Cycles: mode7=          118,034
+	Cycles: lookup_map=      22,747
+	Cycles: put_sprite=       2,561
+	================================
+	Total =                 193,214
+	Frame Rate = 5.18 fps
 
-32*40=12800
+Each cycle removed from inner X loop saves
+	32*40=1280 cycles

tfv/tfv_flying.s

@@ -919,7 +919,7 @@ odd_branch:
 								;==========
 								;	 28
 
-	clc		; fixed_add(&space_y,&temp,&space_y);	; 2
+	clc		; fixed_add(&space_y,&temp,&space_y);		; 2
 	lda	SPACEY_F						; 3
 ;	adc	TEMP_F							;
 	adc	RESULT+1						; 3
@@ -931,23 +931,12 @@ odd_branch:
 
 	; brk	; space_y = f7:04
 
-	lda	#0							; 2
+	ldx	#0	; was SCREEN_X					; 2
-	sta	SCREEN_X						; 3
 								;==========
-								;	 25
+								;	 22
 screenx_loop:
 
-	; cache color and return if same as last time
+
-	lda	SPACEY_I						; 3
-	cmp	LAST_SPACEY_I						; 3
-	bne	nomatch							; 2nt/3
-	lda	SPACEX_I						; 3
-	cmp	LAST_SPACEX_I						; 3
-	bne	nomatch							; 2nt/3
-	lda	LAST_MAP_COLOR						; 3
-	jmp	match							; 3
-								;===========
-								;	22
 nomatch:
 	; do a full lookup, takes much longer
 	jsr	lookup_map		; get color in A		; 6
@@ -1000,13 +989,29 @@ dyi_label:
 	adc	#0							; 2
 	sta	SPACEY_I						; 3
 
-	inc	SCREEN_X						; 5
+	inx	;inc	SCREEN_X					; 2
-	lda	SCREEN_X						; 3
+	cpx	#40			; LOWRES width			; 2
-	cmp	#40			; LOWRES width			; 2
+	beq	done_screenx_loop					; 2nt/3
-	bne	screenx_loop						; 2nt/3
 								;=============
-								;	49
+								;	43
 
+
+	; cache color and return if same as last time
+	lda	SPACEY_I						; 3
+spacey_label:
+	cmp	#0	; LAST_SPACEY_I				; 2
+	bne	nomatch							; 2nt/3
+	lda	SPACEX_I						; 3
+spacex_label:
+	cmp	#0	; LAST_SPACEX_I				; 2
+	bne	nomatch							; 2nt/3
+map_color_label:
+	lda	#0	; LAST_MAP_COLOR			; 2
+	jmp	match							; 3
+								;===========
+								; max 19
+
+done_screenx_loop:
 	inc	SCREEN_Y						; 5
 	lda	SCREEN_Y						; 3
 	cmp	#40			; LOWRES height			; 2
@@ -1041,13 +1046,13 @@ lookup_map:
 ;nomatch:
 	lda	SPACEX_I						; 3
 ;nomatch2:
-	sta	LAST_SPACEX_I						; 3
+	sta	spacex_label+1	; LAST_SPACEX_I			; 4
 	and	#CONST_MAP_MASK_X					; 2
 	sta	SPACEX_I						; 3
 	tay								; 2
 
 	lda	SPACEY_I						; 3
-	sta	LAST_SPACEY_I						; 3
+	sta	spacey_label+1	; LAST_SPACEY_I			; 4
 	and	#CONST_MAP_MASK_Y	; wrap to 64x64 grid		; 2
 	sta	SPACEY_I						; 3
 
@@ -1062,7 +1067,7 @@ lookup_map:
 	; SPACEX_I is in y
 	cpy	#$8							; 2
 								;============
-								;	 37
+								;	 39
 
 	bcs	ocean_color		; bgt 8				;^2nt/3
 	ldy	SPACEY_I						; 3
@@ -1080,7 +1085,7 @@ ocean_color:
 	lda	water_map,Y		; the color of the sea		; 4
 
 update_cache:
-	sta	LAST_MAP_COLOR						; 3
+	sta	map_color_label+1	; self-modifying		; 4
 	rts								; 6
 
 flying_map: