diff --git a/gr-sim/gr-sim.c b/gr-sim/gr-sim.c
index 9adf358a..72e8c505 100644
--- a/gr-sim/gr-sim.c
+++ b/gr-sim/gr-sim.c
@@ -1340,7 +1340,8 @@ hlin_loop:
 	return 0;
 }
 
-int hlin_double(int page, int x1, int x2, int at) {
+int hlin_setup(int page, int x1, int x2, int at) {
+
 	// page, y, V2, A
 	a=at;
 	y=at/2;
@@ -1351,6 +1352,15 @@ int hlin_double(int page, int x1, int x2, int at) {
 	ram[GBASH]+=(page);
 
 	ram[GBASL]+=x1;
+
+	return 0;
+
+}
+
+int hlin_double(int page, int x1, int x2, int at) {
+
+	hlin_setup(page,x1,x2,at);
+
 	hlin_double_continue(x2-x1+1);
 
 	return 0;
diff --git a/gr-sim/gr-sim.h b/gr-sim/gr-sim.h
index b7e0c8e9..790672f8 100644
--- a/gr-sim/gr-sim.h
+++ b/gr-sim/gr-sim.h
@@ -24,6 +24,7 @@ void basic_print(char *string);
 void basic_inverse(void);
 void basic_normal(void);
 int hlin(int page, int x1, int x2, int at);
+int hlin_setup(int page, int x1, int x2, int at);
 int hlin_continue(int width);
 int hlin_double_continue(int width);
 int hlin_double(int page, int x1, int x2, int at);
diff --git a/gr-sim/tfv_flying_6502.c b/gr-sim/tfv_flying_6502.c
index c74fca56..7451f06b 100644
--- a/gr-sim/tfv_flying_6502.c
+++ b/gr-sim/tfv_flying_6502.c
@@ -183,6 +183,17 @@ static unsigned char horizontal_lookup[7][16] = {
 };
 
 
+struct cycle_counts {
+	int flying;
+	int getkey;
+	int page_flip;
+	int mode7;
+	int multiply;
+	int lookup_map;
+	int put_sprite;
+} cycles;
+
+
 static void fixed_add(unsigned char x_i,unsigned char x_f,
 			unsigned char y_i, unsigned char y_f,
 			unsigned char *z_i, unsigned char *z_f) {
@@ -211,14 +222,23 @@ static void fixed_mul(unsigned char x_i, unsigned char x_f,
 	int result3;
 	int result2,result1,result0;
 	int aa,xx,cc=0,cc2,yy;
-	int negate=0;
+	unsigned char negate;
 
 	num1h=x_i;
 	num1l=x_f;
 
-	if (!(num1h&0x80)) goto check_num2;	// bpl check_num2
+	negate=0;				// lda	#0		2
+						// sta	NEGATE		4
 
-	negate++;				// inc negate
+						// lda	NUM1H		4
+
+						cycles.multiply+=13;
+
+
+	if (!(num1h&0x80)) goto check_num2;	// bpl check_num2	2nt/3
+						cycles.multiply--;
+
+	negate++;				// inc NEGATE		6
 
 	num1l=~num1l;
 	num1h=~num1h;
@@ -232,13 +252,15 @@ static void fixed_mul(unsigned char x_i, unsigned char x_f,
 
 	num1l&=0xff;
 	num1h&=0xff;
+						// total=26
+						cycles.multiply+=26;
 check_num2:
 
 	num2h=y_i;
 	num2l=y_f;
-
+						cycles.multiply+=7;
 	if (!(num2h&0x80)) goto unsigned_multiply;
-
+						cycles.multiply--;
 	negate++;
 
 	num2l=~num2l;
@@ -253,7 +275,7 @@ check_num2:
 
 	num2l&=0xff;
 	num2h&=0xff;
-
+						cycles.multiply+=30;
 unsigned_multiply:
 
 //	if (debug) {
@@ -266,7 +288,8 @@ unsigned_multiply:
 	aa=0;		// lda #0 (sz)
 	result2=aa;	// sta result+2
 	xx=16;		// ldx #16 (sz)
-label_l1:
+						cycles.multiply+=8;
+multiply_mainloop:
 	cc=(num2h&1);	//lsr NUM2+1 (szc)
 	num2h>>=1;
 	num2h&=0x7f;
@@ -281,8 +304,9 @@ label_l1:
 
 	num2l|=(cc<<7);
 	cc=cc2;
-
-	if (cc==0) goto label_l2;	// bcc L2
+						cycles.multiply+=15;
+	if (cc==0) goto shift_output;	// bcc L2
+						cycles.multiply--;
 
 	yy=aa;				// tay (sz)
 	cc=0;				// clc
@@ -295,8 +319,8 @@ label_l1:
 	aa=aa+cc+num1h;			// adc NUM1+1
 	cc=!!(aa&0x100);
 	aa=aa&0xff;
-
-label_l2:
+						cycles.multiply+=22;
+shift_output:
 	cc2=aa&1;
 	aa=aa>>1;
 	aa&=0x7f;
@@ -322,8 +346,9 @@ label_l2:
 	cc=cc2;		// ror result+0
 
 	xx--;				// dex
-	if (xx!=0) goto label_l1;	// bne L1
-
+						cycles.multiply+=25;
+	if (xx!=0) goto multiply_mainloop;	// bne L1
+						cycles.multiply--;
 	result3=aa&0xff;		// sta result+3
 
 
@@ -331,10 +356,10 @@ label_l2:
 //		printf("RAW RESULT = %02x:%02x:%02x:%02x\n",
 //			result3&0xff,result2&0xff,result1&0xff,result0&0xff);
 //	}
-
+						cycles.multiply+=13;
 	if (negate&1) {
 //		printf("NEGATING!\n");
-
+						cycles.multiply--;
 		cc=0;
 
 		aa=0;
@@ -356,7 +381,7 @@ label_l2:
 		aa-=result3+cc;
 		cc=!!(aa&0x100);
 		result3=aa;
-
+							cycles.multiply+=50;
 	}
 
 	*z_i=result2&0xff;
@@ -385,7 +410,7 @@ label_l2:
 //	a2=(result3<<24)|(result2<<16)|(result1<<8)|result0;
 //	printf("%d * %d = %d (0x%x)\n",s1,s2,a2,a2);
 
-
+						cycles.multiply+=6;
 	return;
 
 }
@@ -400,14 +425,18 @@ void draw_background_mode7(void) {
 	/* Draw Sky */
 	/* Originally wanted to be fancy and have sun too, but no */
 	color_equals(COLOR_MEDIUMBLUE);
+						cycles.mode7+=10;
+
 	for(ram[SCREEN_Y]=0;ram[SCREEN_Y]<6;ram[SCREEN_Y]+=2) {
 		hlin_double(ram[DRAW_PAGE], 0, 40, ram[SCREEN_Y]);
 	}
-
+						cycles.mode7+=(63+(16*40)+23)*5;
 	/* Draw hazy horizon */
 	color_equals(COLOR_GREY);
 	hlin_double(ram[DRAW_PAGE], 0, 40, 6);
+						cycles.mode7+=14+63+(16*40);
 
+						cycles.mode7+=28;
 	fixed_mul(ram[SPACEZ_I],ram[SPACEZ_F],
 		CONST_BETA_I,CONST_BETA_F,
 		&ram[FACTOR_I],&ram[FACTOR_F]);
@@ -423,15 +452,21 @@ void draw_background_mode7(void) {
 //		fixed_to_double(ram[SPACEZ_I],ram[SPACEZ_F],),
 //		fixed_to_double(&BETA),
 //		fixed_to_double(ram[FACTOR_I],ram[FACTOR_F]));
+							cycles.mode7+=22;
 
 	for (ram[SCREEN_Y] = 8; ram[SCREEN_Y] < LOWRES_H; ram[SCREEN_Y]+=2) {
 
+		y=0;
+		hlin_setup(ram[DRAW_PAGE],y,0,ram[SCREEN_Y]);
+							cycles.mode7+=43;
+
 		// then calculate the horizontal scale, or the distance between
 		// space points on this horizontal line
 		ram[HORIZ_SCALE_I]=0;
 		ram[HORIZ_SCALE_F]=
 			horizontal_lookup[ram[SPACEZ_I]&0xf]
 						[(ram[SCREEN_Y]-8)/2];
+							cycles.mode7+=44;
 
 		if (!displayed) {
 			printf("HORIZ_SCALE %x %x\n",
@@ -442,7 +477,7 @@ void draw_background_mode7(void) {
 		fixed_mul(ram[HORIZ_SCALE_I],ram[HORIZ_SCALE_F],
 			CONST_SCALE_I,CONST_SCALE_F,
 			&ram[DISTANCE_I],&ram[DISTANCE_F]);
-
+							cycles.mode7+=44;
 		if (!displayed) {
 			printf("DISTANCE %x:%x\n",ram[DISTANCE_I],ram[DISTANCE_F]);
 		}
@@ -451,10 +486,12 @@ void draw_background_mode7(void) {
 		// through all points on this line
 		ram[DX_I]=fixed_sin[(ram[ANGLE]+8)&0xf].i;	// -sin()
 		ram[DX_F]=fixed_sin[(ram[ANGLE]+8)&0xf].f;	// -sin()
+							cycles.mode7+=29;
+
 		fixed_mul(ram[DX_I],ram[DX_F],
 			ram[HORIZ_SCALE_I],ram[HORIZ_SCALE_F],
 			&ram[DX_I],&ram[DX_F]);
-
+							cycles.mode7+=48;
 		if (!displayed) {
 			printf("DX %x:%x\n",ram[DX_I],ram[DX_F]);
 		}
@@ -462,10 +499,11 @@ void draw_background_mode7(void) {
 
 		ram[DY_I]=fixed_sin[(ram[ANGLE]+4)&0xf].i;	// cos()
 		ram[DY_F]=fixed_sin[(ram[ANGLE]+4)&0xf].f;	// cos()
+							cycles.mode7+=29;
 		fixed_mul(ram[DY_I],ram[DY_F],
 			ram[HORIZ_SCALE_I],ram[HORIZ_SCALE_F],
 			&ram[DY_I],&ram[DY_F]);
-
+							cycles.mode7+=48;
 		if (!displayed) {
 			printf("DY %x:%x\n",ram[DY_I],ram[DY_F]);
 		}
@@ -474,48 +512,62 @@ void draw_background_mode7(void) {
 		fixed_add(ram[DISTANCE_I],ram[DISTANCE_F],
 			ram[FACTOR_I],ram[FACTOR_F],
 			&ram[SPACEX_I],&ram[SPACEX_F]);
+		fixed_add(ram[DISTANCE_I],ram[DISTANCE_F],
+			ram[FACTOR_I],ram[FACTOR_F],
+			&ram[SPACEY_I],&ram[SPACEY_F]);
+							cycles.mode7+=26;
+
 		ram[TEMP_I]=fixed_sin[(ram[ANGLE]+4)&0xf].i; // cos
 		ram[TEMP_F]=fixed_sin[(ram[ANGLE]+4)&0xf].f; // cos
+							cycles.mode7+=29;
+
 		fixed_mul(ram[SPACEX_I],ram[SPACEX_F],
 			ram[TEMP_I],ram[TEMP_F],
 			&ram[SPACEX_I],&ram[SPACEX_F]);
+							cycles.mode7+=48;
+
 		fixed_add(ram[SPACEX_I],ram[SPACEX_F],
 			ram[CX_I],ram[CX_F],
 			&ram[SPACEX_I],&ram[SPACEX_F]);
 		ram[TEMP_I]=0xec;	// -20 (LOWRES_W/2)
 		ram[TEMP_F]=0;
+							cycles.mode7+=30;
+
 		fixed_mul(ram[TEMP_I],ram[TEMP_F],
 			ram[DX_I],ram[DX_F],
 			&ram[TEMP_I],&ram[TEMP_F]);
+							cycles.mode7+=48;
+
 		fixed_add(ram[SPACEX_I],ram[SPACEX_F],
 			ram[TEMP_I],ram[TEMP_F],
 			&ram[SPACEX_I],&ram[SPACEX_F]);
-
+							cycles.mode7+=20;
 		if (!displayed) {
 			printf("SPACEX! %x:%x\n",
 			ram[SPACEX_I],ram[SPACEX_F]);
 		}
 
-		fixed_add(ram[DISTANCE_I],ram[DISTANCE_F],
-			ram[FACTOR_I],ram[FACTOR_F],
-			&ram[SPACEY_I],&ram[SPACEY_F]);
 		ram[TEMP_I]=fixed_sin[ram[ANGLE]&0xf].i;
 		ram[TEMP_F]=fixed_sin[ram[ANGLE]&0xf].f;
+							cycles.mode7+=25;
 		fixed_mul(ram[SPACEY_I],ram[SPACEY_F],
 			ram[TEMP_I],ram[TEMP_F],
 			&ram[SPACEY_I],&ram[SPACEY_F]);
+							cycles.mode7+=48;
 		fixed_add(ram[SPACEY_I],ram[SPACEY_F],
 			ram[CY_I],ram[CY_F],
 			&ram[SPACEY_I],&ram[SPACEY_F]);
 		ram[TEMP_I]=0xec;	// -20 (LOWRES_W/2)
 		ram[TEMP_F]=0;
+							cycles.mode7+=30;
 		fixed_mul(ram[TEMP_I],ram[TEMP_F],
 			ram[DY_I],ram[DY_F],
 			&ram[TEMP_I],&ram[TEMP_F]);
+							cycles.mode7+=48;
 		fixed_add(ram[SPACEY_I],ram[SPACEY_F],
 			ram[TEMP_I],ram[TEMP_F],
 			&ram[SPACEY_I],&ram[SPACEY_F]);
-
+							cycles.mode7+=25;
 		if (!displayed) {
 			printf("SPACEY! %x:%x\n",ram[SPACEY_I],ram[SPACEY_F]);
 		}
@@ -533,8 +585,8 @@ void draw_background_mode7(void) {
 				if (map_color==COLOR_DARKBLUE) ram[OVER_WATER]=1;
 			}
 
-			hlin_double(ram[DRAW_PAGE], ram[SCREEN_X], ram[SCREEN_X]+1,
-				ram[SCREEN_Y]);
+			hlin_double_continue(1);
+							cycles.mode7+=42;
 
 			// advance to the next position in space
 			fixed_add(ram[SPACEX_I],ram[SPACEX_F],
@@ -543,19 +595,17 @@ void draw_background_mode7(void) {
 			fixed_add(ram[SPACEY_I],ram[SPACEY_F],
 				ram[DY_I],ram[DY_F],
 				&ram[SPACEY_I],&ram[SPACEY_F]);
+							cycles.mode7+=53;
 		}
+							cycles.mode7+=17;
 	}
 	displayed=1;
+							cycles.mode7+=6;
 }
 
 
 
 
-struct cycle_counts {
-	int flying;
-	int getkey;
-	int page_flip;
-} cycles;
 
 static int iterations=0;
 
@@ -775,10 +825,16 @@ int flying(void) {
 		if (iterations==100) {
 			int total_cycles;
 			total_cycles=cycles.flying+cycles.getkey+
-				cycles.page_flip;
+				cycles.page_flip+cycles.multiply+
+				cycles.mode7+cycles.lookup_map+
+				cycles.put_sprite;
 			printf("Cycles: flying=%d\n",cycles.flying);
 			printf("Cycles: getkey=%d\n",cycles.getkey);
 			printf("Cycles: page_flip=%d\n",cycles.page_flip);
+			printf("Cycles: multiply=%d\n",cycles.multiply);
+			printf("Cycles: mode7=%d\n",cycles.mode7);
+			printf("Cycles: lookup_map=%d\n",cycles.lookup_map);
+			printf("Cycles: put_sprite=%d\n",cycles.put_sprite);
 			printf("Total = %d\n",total_cycles);
 			printf("Frame Rate = %.2lf fps\n",
 				(1000000.0 / (double)total_cycles));
diff --git a/tfv/TODO b/tfv/TODO
index 34ac21c1..37e54562 100644
--- a/tfv/TODO
+++ b/tfv/TODO
@@ -1,3 +1,9 @@
+mode7 speed fixes:
+	+ Don't draw sky every frame, only if needed
+	+ faster multiply routine
+	+ move multiply vars to zero page
+	+ leave one of multiply results in accumulator at end?
+
 short term:
 
 longer term:
diff --git a/tfv/tfv_flying.s b/tfv/tfv_flying.s
index 3eb501e6..f36b17de 100644
--- a/tfv/tfv_flying.s
+++ b/tfv/tfv_flying.s
@@ -550,66 +550,60 @@ draw_ship:
 draw_background_mode7:
 
 	; Draw Sky
+	; FIXME: the sky never changes?
 
-	lda	#COLOR_BOTH_MEDIUMBLUE	; MEDIUMBLUE color
-	sta	COLOR
+	lda	#COLOR_BOTH_MEDIUMBLUE	; MEDIUMBLUE color		; 2
+	sta	COLOR							; 3
 
-	lda	#0
-	sta	OVER_WATER
+	lda	#0							; 2
+	sta	OVER_WATER						; 3
+								;===========
+								;	 10
 
 sky_loop:				; draw line across screen
-	ldy	#40			; from y=0 to y=6
-	sty	V2
-	ldy	#0
-	pha
-	jsr	hlin_double		; hlin y,V2 at A
-	pla
-	clc
-	adc	#2
-	cmp	#6
-	bne	sky_loop
-
+	ldy	#40			; from y=0 to y=6		; 2
+	sty	V2							; 3
+	ldy	#0							; 2
+	pha								; 3
+	jsr	hlin_double		; hlin y,V2 at A	; 63+(X*16)
+	pla								; 4
+	clc								; 2
+	adc	#2							; 2
+	cmp	#6							; 2
+	bne	sky_loop						; 3/2nt
+								;=============
+								; (23+63+(X*16))*5
 	; Draw Horizon
 
-	lda	#COLOR_BOTH_GREY	; Horizon is Grey
-	sta	COLOR
-	lda	#6			; draw single line at 6/7
-	ldy	#40
-	sty	V2			; hlin	Y,V2 at A
-	ldy	#0
-	jsr	hlin_double		; hlin	0,40 at 6
-
+	lda	#COLOR_BOTH_GREY	; Horizon is Grey		; 2
+	sta	COLOR							; 3
+	lda	#6			; draw single line at 6/7	; 2
+	ldy	#40							; 2
+	sty	V2			; hlin	Y,V2 at A		; 3
+	ldy	#0							; 2
+	jsr	hlin_double		; hlin	0,40 at 6	; 63+(X*16)
+								;===========
+								; 63+(X*16)+14
 	; fixed_mul(&space_z,&BETA,&factor);
 
-	lda	SPACEZ_I
-	sta	NUM1H
-	lda	SPACEZ_F
-	sta	NUM1L
+	lda	SPACEZ_I						; 3
+	sta	NUM1H							; 3
+	lda	SPACEZ_F						; 3
+	sta	NUM1L							; 3
 
-	lda	#$ff	; BETA_I
-	sta	NUM2H
-	lda	#$80	; BETA_F
-	sta	NUM2L
+	lda	#$ff	; BETA_I					; 2
+	sta	NUM2H							; 3
+	lda	#$80	; BETA_F					; 2
+	sta	NUM2L							; 3
 
-;; TEST
-;;	lda	#$0
-;;	sta	NUM1H
-;;	lda	#$2
-;;	sta	NUM1L
+	jsr	multiply						; 6
+								;===========
+								;        28
 
-;;	lda	#$0
-;;	sta	NUM2H
-;;	lda	#$3
-;;	sta	NUM2L
-
-
-
-	jsr	multiply
-
-	lda	RESULT+2
-	sta	FACTOR_I
-	lda	RESULT+1
-	sta	FACTOR_F
+	lda	RESULT+2						; 4
+	sta	FACTOR_I						; 4
+	lda	RESULT+1						; 4
+	sta	FACTOR_F						; 4
 
 	;; SPACEZ=78  * ff80 = FACTOR=66
 
@@ -617,329 +611,356 @@ sky_loop:				; draw line across screen
 	;; GOOD 4 80 * ffffffff 80 = fffffffd c0
 	;; BAD  4 80 * ffffffff 80 = 42 40
 
-	lda	#8
-	sta	SCREEN_Y
-
+	lda	#8							; 2
+	sta	SCREEN_Y						; 4
+								;=============
+								;	 22
 screeny_loop:
-	ldy	#0
-	jsr	hlin_setup		; y-coord in a, x-coord in y
+	ldy	#0							; 2
+	jsr	hlin_setup		; y-coord in a, x-coord in y	; 41
 					; sets up GBASL/GBASH
+								;=============
+								;	 43
 
-	lda	#0			; horizontal_scale.i = 0
-	sta	HORIZ_SCALE_I
+	lda	#0			; horizontal_scale.i = 0	; 2
+	sta	HORIZ_SCALE_I						; 3
 
 	;horizontal_scale.f=
 	;	horizontal_lookup[space_z.i&0xf][(screen_y-8)/2];
 
-	lda	SPACEZ_I
-	and	#$f
-	asl
-	asl
-	asl
-	asl
-	sta	TEMP_I
-
-	sec
-	lda	SCREEN_Y
-	sbc	#8
-	lsr
-	clc
-	adc	TEMP_I
-	tay
-	lda	horizontal_lookup,Y
-	sta	HORIZ_SCALE_F
+	lda	SPACEZ_I						; 3
+	and	#$f							; 2
+	asl								; 2
+	asl								; 2
+	asl								; 2
+	asl								; 2
+	sta	TEMP_I							; 3
 
+	sec								; 2
+	lda	SCREEN_Y						; 3
+	sbc	#8							; 2
+	lsr								; 2
+	clc								; 2
+	adc	TEMP_I							; 3
+	tay								; 2
+	lda	horizontal_lookup,Y					; 4
+	sta	HORIZ_SCALE_F						; 3
+								;============
+								;	 44
 	;; brk ASM, horiz_scale = 00:73
 
 	; calculate the distance of the line we are drawing
 	; fixed_mul(&horizontal_scale,&scale,&distance);
-	lda	HORIZ_SCALE_I
-	sta	NUM1H
-	lda	HORIZ_SCALE_F
-	sta	NUM1L
-	lda	#$14	; SCALE_I
-	sta	NUM2H
-	lda	#$00	; SCALE_F
-	sta	NUM2L
-	jsr	multiply
-	lda	RESULT+2
-	sta	DISTANCE_I
-	lda	RESULT+1
-	sta	DISTANCE_F
-
+	lda	HORIZ_SCALE_I						; 3
+	sta	NUM1H							; 4
+	lda	HORIZ_SCALE_F						; 3
+	sta	NUM1L							; 4
+	lda	#$14	; SCALE_I					; 2
+	sta	NUM2H							; 4
+	lda	#$00	; SCALE_F					; 2
+	sta	NUM2L							; 4
+	jsr	multiply						; 6
+	lda	RESULT+2						; 4
+	sta	DISTANCE_I						; 2
+	lda	RESULT+1						; 4
+	sta	DISTANCE_F						; 2
+								;==========
+								;	 44
 	;; brk ASM, distance = 08:fc
 
 	; calculate the dx and dy of points in space when we step
 	; through all points on this line
 
-	lda	ANGLE	; dx.i=fixed_sin[(angle+8)&0xf].i; // -sin()
-	clc
-	adc	#8
-	and	#$f
-	asl
-	tay
-	lda	fixed_sin,Y
-	sta	DX_I
-	iny		; dx.f=fixed_sin[(angle+8)&0xf].f; // -sin()
-	lda	fixed_sin,Y
-	sta	DX_F
+	lda	ANGLE	; dx.i=fixed_sin[(angle+8)&0xf].i; // -sin()	; 3
+	clc								; 2
+	adc	#8							; 2
+	and	#$f							; 2
+	asl								; 2
+	tay								; 2
+	lda	fixed_sin,Y						; 4
+	sta	DX_I							; 3
+	iny		; dx.f=fixed_sin[(angle+8)&0xf].f; // -sin()	; 2
+	lda	fixed_sin,Y						; 4
+	sta	DX_F							; 3
+								;==========
+								;	 29
 
 	; fixed_mul(&dx,&horizontal_scale,&dx);
-	lda	HORIZ_SCALE_I
-	sta	NUM1H
-	lda	HORIZ_SCALE_F
-	sta	NUM1L
-	lda	DX_I
-	sta	NUM2H
-	lda	DX_F
-	sta	NUM2L
-	jsr	multiply
-	lda	RESULT+2
-	sta	DX_I
-	lda	RESULT+1
-	sta	DX_F
-
+	lda	HORIZ_SCALE_I						; 3
+	sta	NUM1H							; 4
+	lda	HORIZ_SCALE_F						; 3
+	sta	NUM1L							; 4
+	lda	DX_I							; 3
+	sta	NUM2H							; 4
+	lda	DX_F							; 3
+	sta	NUM2L							; 4
+	jsr	multiply						; 6
+	lda	RESULT+2						; 4
+	sta	DX_I							; 3
+	lda	RESULT+1						; 4
+	sta	DX_F							; 3
+								;==========
+								;	 48
 	;; ANGLE
 	;; brk ASM, dx = 00:00
 
-	lda	ANGLE		; dy.i=fixed_sin[(angle+4)&0xf].i; // cos()
-	clc
-	adc	#4
-	and	#$f
-	asl
-	tay
-	lda	fixed_sin,Y
-	sta	DY_I
-	iny			; dy.f=fixed_sin[(angle+4)&0xf].f; // cos()
-	lda	fixed_sin,Y
-	sta	DY_F
-
+	lda	ANGLE	; dy.i=fixed_sin[(angle+4)&0xf].i; // cos()	; 3
+	clc								; 2
+	adc	#4							; 2
+	and	#$f							; 2
+	asl								; 2
+	tay								; 2
+	lda	fixed_sin,Y						; 4
+	sta	DY_I							; 3
+	iny		; dy.f=fixed_sin[(angle+4)&0xf].f; // cos()	; 2
+	lda	fixed_sin,Y						; 4
+	sta	DY_F							; 3
+								;==========
+								;	 29
 	; fixed_mul(&dy,&horizontal_scale,&dy);
-	lda	HORIZ_SCALE_I
-	sta	NUM1H
-	lda	HORIZ_SCALE_F
-	sta	NUM1L
-	lda	DY_I
-	sta	NUM2H
-	lda	DY_F
-	sta	NUM2L
-	jsr	multiply
-	lda	RESULT+2
-	sta	DY_I
-	lda	RESULT+1
-	sta	DY_F
-
+	lda	HORIZ_SCALE_I						; 3
+	sta	NUM1H							; 4
+	lda	HORIZ_SCALE_F						; 3
+	sta	NUM1L							; 4
+	lda	DY_I							; 3
+	sta	NUM2H							; 4
+	lda	DY_F							; 3
+	sta	NUM2L							; 4
+	jsr	multiply						; 6
+	lda	RESULT+2						; 4
+	sta	DY_I							; 3
+	lda	RESULT+1						; 4
+	sta	DY_F							; 3
+								;==========
+								;	 48
 	;; brk ASM, dy = 00:73
 
 	; calculate the starting position
 
-				; fixed_add(&distance,&factor,&space_x);
-	clc			; fixed_add(&distance,&factor,&space_y);
-	lda	DISTANCE_F
-	adc	FACTOR_F
-	sta	SPACEY_F
-	sta	SPACEX_F
-	lda	DISTANCE_I
-	adc	FACTOR_I
-	sta	SPACEY_I
-	sta	SPACEX_I
-
+			; fixed_add(&distance,&factor,&space_x);
+	clc		; fixed_add(&distance,&factor,&space_y);	; 2
+	lda	DISTANCE_F						; 3
+	adc	FACTOR_F						; 3
+	sta	SPACEY_F						; 3
+	sta	SPACEX_F						; 3
+	lda	DISTANCE_I						; 3
+	adc	FACTOR_I						; 3
+	sta	SPACEY_I						; 3
+	sta	SPACEX_I						; 3
+								;==========
+								;	 26
 	;; brk	space_x = 06:bc
 
-	lda	ANGLE	; fixed_temp.i=fixed_sin[(angle+4)&0xf].i; // cos
-	clc
-	adc	#4
-	and	#$f
-	asl
-	tay
-	lda	fixed_sin,Y
-	sta	TEMP_I
-	iny		; fixed_temp.f=fixed_sin[(angle+4)&0xf].f; // cos
-	lda	fixed_sin,Y
-	sta	TEMP_F
+	lda	ANGLE	; temp.i=fixed_sin[(angle+4)&0xf].i; // cos	; 3
+	clc								; 2
+	adc	#4							; 2
+	and	#$f							; 2
+	asl								; 2
+	tay								; 2
+	lda	fixed_sin,Y						; 4
+	sta	TEMP_I							; 3
+	iny		; temp.f=fixed_sin[(angle+4)&0xf].f; // cos	; 2
+	lda	fixed_sin,Y						; 4
+	sta	TEMP_F							; 3
+								;==========
+								;	 29
 
-	; fixed_mul(&space_x,&fixed_temp,&space_x);
-	lda	SPACEX_I
-	sta	NUM1H
-	lda	SPACEX_F
-	sta	NUM1L
-	lda	TEMP_I
-	sta	NUM2H
-	lda	TEMP_F
-	sta	NUM2L
-	jsr	multiply
-	lda	RESULT+2
-	sta	SPACEX_I
-	lda	RESULT+1
-	sta	SPACEX_F
+	; fixed_mul(&space_x,&temp,&space_x);
+	lda	SPACEX_I						; 3
+	sta	NUM1H							; 4
+	lda	SPACEX_F						; 3
+	sta	NUM1L							; 4
+	lda	TEMP_I							; 3
+	sta	NUM2H							; 4
+	lda	TEMP_F							; 3
+	sta	NUM2L							; 4
+	jsr	multiply						; 6
+	lda	RESULT+2						; 4
+	sta	SPACEX_I						; 3
+	lda	RESULT+1						; 4
+	sta	SPACEX_F						; 3
+								;==========
+								;	 48
+
+	clc			; fixed_add(&space_x,&cx,&space_x);	; 2
+	lda	SPACEX_F						; 3
+	adc	CX_F							; 3
+	sta	SPACEX_F						; 3
+	lda	SPACEX_I						; 3
+	adc	CX_I							; 3
+	sta	SPACEX_I						; 3
+
+	lda	#$ec		; temp.i=0xec;	// -20 (LOWRES_W/2)	; 2
+	sta	TEMP_I							; 3
+	lda	#0		; temp.f=0;				; 2
+	sta	TEMP_F							; 3
+								;==========
+								;	 30
+	; fixed_mul(&temp,&dx,&temp);
+	lda	TEMP_I							; 3
+	sta	NUM1H							; 4
+	lda	TEMP_F							; 3
+	sta	NUM1L							; 4
+	lda	DX_I							; 3
+	sta	NUM2H							; 4
+	lda	DX_F							; 3
+	sta	NUM2L							; 4
+	jsr	multiply						; 6
+	lda	RESULT+2						; 4
+	sta	TEMP_I							; 3
+	lda	RESULT+1						; 4
+	sta	TEMP_F							; 3
+								;==========
+								;	 48
 
 
-	clc			; fixed_add(&space_x,&cx,&space_x);
-	lda	SPACEX_F
-	adc	CX_F
-	sta	SPACEX_F
-	lda	SPACEX_I
-	adc	CX_I
-	sta	SPACEX_I
 
-	lda	#$ec		; fixed_temp.i=0xec;      // -20 (LOWRES_W/2)
-	sta	TEMP_I
-	lda	#0		; fixed_temp.f=0;
-	sta	TEMP_F
-
-	; fixed_mul(&fixed_temp,&dx,&fixed_temp);
-	lda	TEMP_I
-	sta	NUM1H
-	lda	TEMP_F
-	sta	NUM1L
-	lda	DX_I
-	sta	NUM2H
-	lda	DX_F
-	sta	NUM2L
-	jsr	multiply
-	lda	RESULT+2
-	sta	TEMP_I
-	lda	RESULT+1
-	sta	TEMP_F
-
-
-	clc			; fixed_add(&space_x,&fixed_temp,&space_x);
-	lda	SPACEX_F
-	adc	TEMP_F
-	sta	SPACEX_F
-	lda	SPACEX_I
-	adc	TEMP_I
-	sta	SPACEX_I
-
-;;brk	;; brk	space_x = 06:bc
-
-	lda	ANGLE	; fixed_temp.i=fixed_sin[angle&0xf].i;
-	and	#$f
-	asl
-	tay
-	lda	fixed_sin,Y
-	sta	TEMP_I
-	iny		; fixed_temp.f=fixed_sin[angle&0xf].f;
-	lda	fixed_sin,Y
-	sta	TEMP_F
+	clc		; fixed_add(&space_x,&temp,&space_x);		; 2
+	lda	SPACEX_F						; 3
+	adc	TEMP_F							; 3
+	sta	SPACEX_F						; 3
+	lda	SPACEX_I						; 3
+	adc	TEMP_I							; 3
+	sta	SPACEX_I						; 3
+								;==========
+								;	 20
+	; brk	; space_x = 06:bc
 
+	lda	ANGLE	; temp.i=fixed_sin[angle&0xf].i;		; 3
+	and	#$f							; 2
+	asl								; 2
+	tay								; 2
+	lda	fixed_sin,Y						; 4
+	sta	TEMP_I							; 3
+	iny		; fixed_temp.f=fixed_sin[angle&0xf].f;		; 2
+	lda	fixed_sin,Y						; 4
+	sta	TEMP_F							; 3
+								;==========
+								;	 25
 
 	; fixed_mul(&space_y,&fixed_temp,&space_y);
-	lda	SPACEY_I
-	sta	NUM1H
-	lda	SPACEY_F
-	sta	NUM1L
-	lda	TEMP_I
-	sta	NUM2H
-	lda	TEMP_F
-	sta	NUM2L
-	jsr	multiply
-	lda	RESULT+2
-	sta	SPACEY_I
-	lda	RESULT+1
-	sta	SPACEY_F
+	lda	SPACEY_I						; 3
+	sta	NUM1H							; 4
+	lda	SPACEY_F						; 3
+	sta	NUM1L							; 4
+	lda	TEMP_I							; 3
+	sta	NUM2H							; 4
+	lda	TEMP_F							; 3
+	sta	NUM2L							; 4
+	jsr	multiply						; 6
+	lda	RESULT+2						; 4
+	sta	SPACEY_I						; 3
+	lda	RESULT+1						; 4
+	sta	SPACEY_F						; 3
+								;==========
+								;	 48
 
+	clc			; fixed_add(&space_y,&cy,&space_y);	; 2
+	lda	SPACEY_F						; 3
+	adc	CY_F							; 3
+	sta	SPACEY_F						; 3
+	lda	SPACEY_I						; 3
+	adc	CY_I							; 3
+	sta	SPACEY_I						; 3
 
-	clc			; fixed_add(&space_y,&cy,&space_y);
-	lda	SPACEY_F
-	adc	CY_F
-	sta	SPACEY_F
-	lda	SPACEY_I
-	adc	CY_I
-	sta	SPACEY_I
-
-	lda	#$ec		; fixed_temp.i=0xec;      // -20 (LOWRES_W/2)
-	sta	TEMP_I
-	lda	#0		; fixed_temp.f=0;
-	sta	TEMP_F
-
+	lda	#$ec		; temp.i=0xec;	// -20 (LOWRES_W/2)	; 2
+	sta	TEMP_I							; 3
+	lda	#0		; temp.f=0;				; 2
+	sta	TEMP_F							; 3
+								;==========
+								;	 30
 	; fixed_mul(&fixed_temp,&dy,&fixed_temp);
-	lda	TEMP_I
-	sta	NUM1H
-	lda	TEMP_F
-	sta	NUM1L
-	lda	DY_I
-	sta	NUM2H
-	lda	DY_F
-	sta	NUM2L
-	jsr	multiply
-	lda	RESULT+2
-	sta	TEMP_I
-	lda	RESULT+1
-	sta	TEMP_F
+	lda	TEMP_I							; 3
+	sta	NUM1H							; 4
+	lda	TEMP_F							; 3
+	sta	NUM1L							; 4
+	lda	DY_I							; 3
+	sta	NUM2H							; 4
+	lda	DY_F							; 3
+	sta	NUM2L							; 4
+	jsr	multiply						; 6
+	lda	RESULT+2						; 4
+	sta	TEMP_I							; 3
+	lda	RESULT+1						; 4
+	sta	TEMP_F							; 3
+								;==========
+								;	 48
 
+	clc		; fixed_add(&space_y,&fixed_temp,&space_y);	; 2
+	lda	SPACEY_F						; 3
+	adc	TEMP_F							; 3
+	sta	SPACEY_F						; 3
+	lda	SPACEY_I						; 3
+	adc	TEMP_I							; 3
+	sta	SPACEY_I						; 3
 
-	clc			; fixed_add(&space_y,&fixed_temp,&space_y);
-	lda	SPACEY_F
-	adc	TEMP_F
-	sta	SPACEY_F
-	lda	SPACEY_I
-	adc	TEMP_I
-	sta	SPACEY_I
+	; brk	; space_y = f7:04
 
-;;brk	;; brk	space_y = f7:04
-
-	lda	#0
-	sta	SCREEN_X
+	lda	#0							; 2
+	sta	SCREEN_X						; 3
+								;==========
+								;	 25
 screenx_loop:
 
-	jsr	lookup_map		; get color in A
+	jsr	lookup_map		; get color in A		; 6
 
-	ldy	#0
-	sta	(GBASL),Y		; plot double height
-	inc	GBASL			; point to next pixel
+	ldy	#0							; 2
+	sta	(GBASL),Y		; plot double height		; 6
+	inc	GBASL			; point to next pixel		; 5
 
 	; Check if over water
-	cmp	#$22			; see if dark blue
-	bne	not_watery
+	cmp	#$22			; see if dark blue		; 2
+	bne	not_watery						; 2nt/3
 
-	lda	SCREEN_Y		; only check pixel in middle of screen
-	cmp	#38
-	bne	not_watery
+	lda	SCREEN_Y	; only check pixel in middle of screen	; 3
+	cmp	#38							; 2
+	bne	not_watery						; 2nt/3
 
-	lda	SCREEN_X		; only check pixel in middle of screen
-	cmp	#20
-	bne	not_watery
-
-	lda	#$1			; set over water
-	sta	OVER_WATER
+	lda	SCREEN_X	; only check pixel in middle of screen	; 3
+	cmp	#20							; 2
+	bne	not_watery						; 2nt/3
 
+	lda	#$1		; set over water			; 2
+	sta	OVER_WATER						; 3
+								;============
+								;	 42
 not_watery:
 	; advance to the next position in space
 
-	clc			; fixed_add(&space_x,&dx,&space_x);
-	lda	SPACEX_F
-	adc	DX_F
-	sta	SPACEX_F
-	lda	SPACEX_I
-	adc	DX_I
-	sta	SPACEX_I
+	clc			; fixed_add(&space_x,&dx,&space_x);	; 2
+	lda	SPACEX_F						; 3
+	adc	DX_F							; 3
+	sta	SPACEX_F						; 3
+	lda	SPACEX_I						; 3
+	adc	DX_I							; 3
+	sta	SPACEX_I						; 3
 
-	clc			; fixed_add(&space_y,&dy,&space_y);
-	lda	SPACEY_F
-	adc	DY_F
-	sta	SPACEY_F
-	lda	SPACEY_I
-	adc	DY_I
-	sta	SPACEY_I
+	clc			; fixed_add(&space_y,&dy,&space_y);	; 2
+	lda	SPACEY_F						; 3
+	adc	DY_F							; 3
+	sta	SPACEY_F						; 3
+	lda	SPACEY_I						; 3
+	adc	DY_I							; 3
+	sta	SPACEY_I						; 3
 
+	inc	SCREEN_X						; 5
+	lda	SCREEN_X						; 3
+	cmp	#40			; LOWRES width			; 2
+	bne	screenx_loop						; 2nt/3
+								;=============
+								;	53
 
-
-	inc	SCREEN_X
-	lda	SCREEN_X
-	cmp	#40			; LOWRES width
-	bne	screenx_loop
-
-
-	lda	SCREEN_Y
-	clc
-	adc	#2
-	sta	SCREEN_Y
-	cmp	#40			; LOWRES height
-	beq	done_screeny
-	jmp	screeny_loop
+	lda	SCREEN_Y						; 3
+	clc								; 2
+	adc	#2							; 2
+	sta	SCREEN_Y						; 3
+	cmp	#40			; LOWRES height			; 2
+	beq	done_screeny						; 2nt/3
+	jmp	screeny_loop						; 3
+								;=============
+								;	 17
 done_screeny:
-	rts
+	rts								; 6
 
 
 	;====================
diff --git a/tfv/tfv_multiply.s b/tfv/tfv_multiply.s
index a1333002..10e80a36 100644
--- a/tfv/tfv_multiply.s
+++ b/tfv/tfv_multiply.s
@@ -15,96 +15,96 @@ NEGATE:	.byte 0
 
 multiply:
 
-	lda	#$0
-	sta	NEGATE
+	lda	#$0							; 2
+	sta	NEGATE							; 4
 
 	; Handle Signed
-	lda	NUM1H
-	bpl	check_num2
+	lda	NUM1H							; 4
+	bpl	check_num2						; 2nt/3
 
-	inc	NEGATE
+	inc	NEGATE							; 4
 
-	clc
-	lda	NUM1L
-	eor	#$ff
-	adc	#$1
-	sta	NUM1L
+	clc		; 2s-complement NUM1H/NUM1L			; 2
+	lda	NUM1L							; 4
+	eor	#$ff							; 2
+	adc	#$1							; 2
+	sta	NUM1L							; 4
 
-	lda	NUM1H
-	eor	#$ff
-	adc	#$0
-	sta	NUM1H
+	lda	NUM1H							; 4
+	eor	#$ff							; 2
+	adc	#$0							; 2
+	sta	NUM1H							; 4
 
 check_num2:
-	lda	NUM2H
-	bpl	unsigned_multiply
+	lda	NUM2H							; 4
+	bpl	unsigned_multiply					; 2nt/3
 
-	inc	NEGATE
+	inc	NEGATE							; 4
 
-	clc
-	lda	NUM2L
-	eor	#$ff
-	adc	#$1
-	sta	NUM2L
+	clc								; 2
+	lda	NUM2L							; 4
+	eor	#$ff							; 2
+	adc	#$1							; 2
+	sta	NUM2L							; 4
 
-	lda	NUM2H
-	eor	#$ff
-	adc	#$0
-	sta	NUM2H
+	lda	NUM2H							; 4
+	eor	#$ff							; 2
+	adc	#$0							; 2
+	sta	NUM2H							; 4
 
 unsigned_multiply:
 
-	lda	#0		; Initialize RESULT to 0
-	sta 	RESULT+2
-	ldx	#16		; 16x16 multiply
+	lda	#0		; Initialize RESULT to 0		; 2
+	sta 	RESULT+2						; 4
+	ldx	#16		; 16x16 multiply			; 2
 multiply_mainloop:
-	lsr	NUM2H		; Shift right 16-bit NUM2
-	ror	NUM2L		; low bit goes into carry
-	bcc	shift_output	; 0 or 1?
-	tay			; If 1, add NUM1 (hi byte of RESULT is in A)
-	clc
-	lda	NUM1L
-	adc	RESULT+2
-	sta	RESULT+2
-	tya
-	adc	NUM1H
+	lsr	NUM2H		; Shift right 16-bit NUM2		; 6
+	ror	NUM2L		; low bit goes into carry		; 6
+	bcc	shift_output	; 0 or 1?				; 2nt/3
+	tay			; If 1, add NUM1 (hi byte RESULT in A)	; 2
+	clc								; 2
+	lda	NUM1L							; 4
+	adc	RESULT+2						; 4
+	sta	RESULT+2						; 4
+	tya								; 2
+	adc	NUM1H							; 4
 shift_output:
-	ror	A		; "Stairstep" shift
-	ror	RESULT+2
-	ror	RESULT+1
-	ror	RESULT
-	dex
-	bne	multiply_mainloop
-	sta	RESULT+3
+	ror	A		; "Stairstep" shift			; 2
+	ror	RESULT+2						; 6
+	ror	RESULT+1						; 6
+	ror	RESULT							; 6
+	dex								; 2
+	bne	multiply_mainloop					; 2nt/3
+	sta	RESULT+3						; 4
 
 	;; Negate if necessary
 
-	lda	NEGATE
-	and	#$1
-	beq	positive
+	lda	NEGATE							; 4
+	and	#$1							; 2
+	beq	positive						; 2nt/3
 
-	clc
-	lda	RESULT+0
-	eor	#$ff
-	adc	#$1
-	sta	RESULT+0
+	clc								; 2
+	lda	RESULT+0						; 4
+	eor	#$ff							; 2
+	adc	#$1							; 2
+	sta	RESULT+0						; 4
 
-	lda	RESULT+1
-	eor	#$ff
-	adc	#$0
-	sta	RESULT+1
+	lda	RESULT+1						; 4
+	eor	#$ff							; 2
+	adc	#$0							; 2
+	sta	RESULT+1						; 4
 
-	lda	RESULT+2
-	eor	#$ff
-	adc	#$0
-	sta	RESULT+2
+	lda	RESULT+2						; 4
+	eor	#$ff							; 2
+	adc	#$0							; 2
+	sta	RESULT+2						; 4
 
-	lda	RESULT+3
-	eor	#$ff
-	adc	#$0
-	sta	RESULT+3
+	lda	RESULT+3						; 4
+	eor	#$ff							; 2
+	adc	#$0							; 2
+	sta	RESULT+3						; 4
 
 positive:
 
-	rts
+	rts								; 6
 
diff --git a/tfv/tfv_utils.s b/tfv/tfv_utils.s
index 79240d5e..73423f23 100644
--- a/tfv/tfv_utils.s
+++ b/tfv/tfv_utils.s
@@ -579,19 +579,20 @@ vlin_too_slow:
 	; put address in GBASL/GBASH
 	; Ycoord in A, Xcoord in Y
 hlin_setup:
-	sty	TEMPY
-	tay			; y=A
-	lda	gr_offsets,Y	; lookup low-res memory address
-	clc
-	adc	TEMPY
-	sta	GBASL
-	iny
-
-	lda	gr_offsets,Y
-	adc	DRAW_PAGE	; add in draw page offset
-	sta	GBASH
-	rts
+	sty	TEMPY							; 3
+	tay			; y=A					; 2
+	lda	gr_offsets,Y	; lookup low-res memory address		; 4
+	clc								; 2
+	adc	TEMPY							; 3
+	sta	GBASL							; 3
+	iny								; 2
 
+	lda	gr_offsets,Y						; 4
+	adc	DRAW_PAGE	; add in draw page offset		; 3
+	sta	GBASH							; 3
+	rts								; 6
+								;===========
+								;	35
 	;================================
 	; hlin_double:
 	;================================
@@ -601,18 +602,16 @@ hlin_setup:
 hlin_double:
 ;int hlin_double(int page, int x1, int x2, int at) {
 
-	jsr	hlin_setup
+	jsr	hlin_setup						; 41
 
-	sec
-	lda	V2
-	sbc	TEMPY
+	sec								; 2
+	lda	V2							; 3
+	sbc	TEMPY							; 3
 
-	tax
-	inx
-
-;	jsr	hlin_double_continue
-
-;	rts
+	tax								; 2
+	inx								; 2
+								;===========
+								;	53
 	; fallthrough
 
 	;=================================
@@ -622,16 +621,17 @@ hlin_double:
 
 hlin_double_continue:
 
+	ldy	#0							; 2
+	lda	COLOR							; 3
 hlin_double_loop:
-	ldy	#0
-	lda	COLOR
-	sta	(GBASL),Y
-	inc	GBASL
-	dex
-	bne	hlin_double_loop
-
-	rts
+	sta	(GBASL),Y						; 6
+	inc	GBASL							; 5
+	dex								; 2
+	bne	hlin_double_loop					; 2nt/3
 
+	rts								; 6
+								;=============
+								; 53+5+X*16+5
 
 	;================================
 	; hlin_single: