small optimizations

apu.s

@@ -2,35 +2,30 @@
 ; Updates the APU registers. x and y are free to use
 ;
 
-.if 0
-; Found this on nesdev bbs by blargg,
-; this can replace the volume table but takes a little more CPU
-ft_get_volume:
-
-	lda var_ch_VolColumn, x
-	lsr a
-	lsr a
-	lsr a
-	sta var_Temp
-	lda var_ch_Volume, x
-	sta var_Temp2
-
-	lda var_Temp				; 4x4 multiplication
+; calculates FLOOR(((var_Temp + 1) * (var_Temp2 + 1) - 1) / 16)
+; 4 effective bits for var_Temp2, each block adds one
+.if .defined(USE_VRC6) || .defined(USE_FDS)
+ft_multiply_volume:		;;; ;; ; 050B
+	lda var_Temp
 	lsr var_Temp2
 	bcs :+
 	lsr a
-:   lsr var_Temp2
+:
+	lsr var_Temp2
 	bcc :+
 	adc var_Temp
-:   lsr a
+:	lsr a
+
 	lsr var_Temp2
 	bcc :+
 	adc var_Temp
-:   lsr a
+:	lsr a
+
 	lsr var_Temp2
 	bcc :+
 	adc var_Temp
-:   lsr a
+:	lsr a
+
 	beq :+
 	rts
 :	lda var_Temp
@@ -478,33 +473,6 @@ ft_update_apu:
 @Return:
 	rts
 
-.if .defined(USE_VRC6) || .defined(USE_FDS)
-ft_multiply_volume:		;;; ;; ; 050B
-	lda var_Temp				; 5x4 multiplication
-	lsr var_Temp2
-	bcs :+
-	lsr a
-:   lsr var_Temp2
-	bcc :+
-	adc var_Temp
-:   lsr a
-	lsr var_Temp2
-	bcc :+
-	adc var_Temp
-:   lsr a
-	lsr var_Temp2
-	bcc :+
-	adc var_Temp
-:   lsr a
-	beq :+
-	rts
-:	lda var_Temp
-	ora var_ch_Volume, x
-	beq :+
-	lda #$01					; Round up to 1
-:	rts
-.endif
-
 ; Lookup tables
 
 ft_duty_table:

init.s

@@ -539,10 +539,12 @@ ft_SkipToRow:
 	sta var_Temp2
 	jmp @Finished
 
+@EffectDispatch:		;;; ;; ;
+	jsr @Effect
 @NoRowDelay:
 	; Read a row
 	lda (var_Temp_Pattern), y
-	bmi @Effect
+	bmi @EffectDispatch
 
 	lda var_ch_DefaultDelay, x
 	cmp #$FF
@@ -611,30 +613,30 @@ ft_SkipToRow:
 	iny									; Command takes two bytes
 @OneByteCommand:						; Command takes one byte
 	iny
-	jmp @NoRowDelay						; A new command or note is immediately following
+	rts									; A new command or note is immediately following
 @EffectDuration:
 	iny
 	lda (var_Temp_Pattern), y
 	iny
 	sta var_ch_DefaultDelay, x
-	jmp @NoRowDelay
+	rts
 @EffectNoDuration:
 	iny
 	lda #$FF
 	sta var_ch_DefaultDelay, x
-	jmp @NoRowDelay
+	rts
 @LoadInstCmd:    ; mult-byte
 	iny
 	lda (var_Temp_Pattern), y
 	iny
 	sta var_Temp3
-	jmp @NoRowDelay
+	rts
 @LoadInst:       ; single byte
 	iny
 	and #$0F
 	asl a
 	sta var_Temp3
-	jmp @NoRowDelay						;;; ;; ; var_ch_NoteDelay remains unaltered
+	rts									;;; ;; ; var_ch_NoteDelay remains unaltered
 
 .else   ; ENABLE_ROW_SKIP
 	rts

instrument.s

@@ -334,15 +334,13 @@ ft_run_sequence:
 	pha
 	lda var_ch_State, x
 	and #STATE_RELEASE				;;; ;; ;
-	bne :+
-	pla
+	bne :++
+:	pla
 	rts								; Return new index
 :	ldy #$02						; Check release point
 	lda (var_Temp_Pointer), y
-	bne :+
-	pla								; Release point not found, loop
-	rts
-:	sta var_Temp					;;; ;; ;
+	beq :--							; Release point not found, loop
+	sta var_Temp					;;; ;; ;
 	dec var_Temp
 	pla								; Release point found, don't loop
 	cmp var_Temp