Added a speed optimization. Makes the code somewhat larger, but is a huge

improvement speedwise.


git-svn-id: svn://svn.cc65.org/cc65/trunk@567 b7a2c559-68d2-44c3-8de9-860c34a00d81

libsrc/common/_printf.s

@@ -6,12 +6,11 @@
 
   	.export	  	__printf
 
-	.import	  	popax, pushax, pusheax, push1, axlong, axulong
+	.import	  	popax, pushax, pusheax, decsp6, push1, axlong, axulong
 	.import		__ctype
 	.import		_ltoa, _ultoa
 	.import		_strlower, _strlen
-	.import		jmpvec
+	.importzp	sp, ptr1, ptr2, tmp1, regbank, sreg
-	.importzp	sp, ptr1, tmp1, regbank, sreg
 
 	.macpack	generic
 
@@ -27,7 +26,8 @@ OutData		= regbank+4 		; Function parameters
 ; Other zero page cells
 
 Base	     	= ptr1
-
+FSave	     	= ptr1
+FCount		= ptr2
 
 .code
 
@@ -60,6 +60,7 @@ Output1:
 	ldx	#>CharArg
 	jsr	pushax
 	jsr  	push1
+CallOutFunc:
 	jmp	(OutFunc)     	; fout (OutData, &CharArg, 1)
 
 ; ----------------------------------------------------------------------------
@@ -282,9 +283,70 @@ Save:	lda	regbank,y
 ; Start parsing the format string
 
 MainLoop:
-	jsr  	GetFormatChar		; Get one char, zero in Y
+	lda	Format			; Remember current format pointer
+	sta	FSave
+	lda	Format+1
+	sta	FSave+1
+
+	ldy	#0  			; Index
+@L1:	lda	(Format),y		; Get next char
+       	beq 	@L2			; Jump on end of string
+	cmp	#'%'			; Format spec?
+	beq	@L2
+   	iny	    			; Bump pointer
+       	bne    	@L1
+	inc	Format+1		; Bump high byte of pointer
+	bne	@L1 			; Branch always
+
+; Found a '%' character or end of string. Update the Format pointer so it is
+; current (points to this character).
+
+@L2:	tya	    		     	; Low byte of offset
+	add	Format
+	sta	Format
+	bcc	@L3
+    	inc	Format+1
+
+; Calculate, how many characters must be output. Beware: This number may
+; be zero. A still contains the low byte of the pointer.
+
+@L3:	sub	FSave
+	sta	FCount
+	lda	Format+1
+	sbc	FSave+1
+	sta	FCount+1
+	ora	FCount 		     	; Is the result zero?
+       	beq	@L4		     	; Jump if yes
+
+; Output the characters that we have until now. To make the call to out
+; faster, build the stack frame by hand (don't use pushax)
+
+	jsr	decsp6		     	; 3 args
+	ldy	#5
+	lda	OutData+1
+	sta	(sp),y
+	dey
+	lda	OutData
+	sta	(sp),y
+	dey
+	lda	FSave+1
+	sta	(sp),y
+	dey
+	lda	FSave
+	sta	(sp),y
+	dey
+	lda	FCount+1
+	sta	(sp),y
+	dey
+	lda	FCount
+	sta	(sp),y
+	jsr	CallOutFunc	     	; Call the output function
+
+; We're back from out(), or we didn't call it. Check for end of string.
+
+@L4:  	jsr  	GetFormatChar	     	; Get one char, zero in Y
     	tax  	      		     	; End of format string reached?
-       	bne	NotDone		      	; Continue of end not reached
+       	bne    	NotDone		     	; End not reached
 
 ; End of format string reached. Restore the zeropage registers and return.