diff --git a/asm_routines/lz4_decode.s b/asm_routines/lz4_decode.s
index 864290a5..57b30257 100644
--- a/asm_routines/lz4_decode.s
+++ b/asm_routines/lz4_decode.s
@@ -131,21 +131,21 @@ done_getsrc:
 	; buildcount
 	;============
 buildcount:
-	ldx	#1
-	stx	count+1
-	cmp	#$0f
+	ldx	#1			; high count starts at 1
+	stx	count+1			; (loops at zero?)
+	cmp	#$0f			; if LITERAL_COUNT < 15, we are done
 	bne	done_buildcount
-minus_buildcount:
-	sta	count
-	jsr	getsrc
-	tax
+buildcount_loop:
+	sta	count			; save LITERAL_COUNT (15)
+	jsr	getsrc			; get the next byte
+	tax				; put in X
 	clc
-	adc	count
-	bcc	skip_buildcount
+	adc	count			; add new byte to old value
+	bcc	bc_8bit_oflow		; if overflow, increment high byte
 	inc	count+1
-skip_buildcount:
-	inx
-	beq	minus_buildcount
+bc_8bit_oflow:
+	inx				; check if read value was 255
+	beq	buildcount_loop		; if it was, keep looping and adding
 done_buildcount:
 	rts
 
@@ -168,14 +168,27 @@ putdst:
 putdst_end:
 	rts
 
-	;================
+	;=============================
 	; docopy
-	;================
+	;=============================
+	; copies ram[count+1]-1:X bytes
+	; from src to dst
 docopy:
+	; working around bug in original code that would loop 256 too
+	; many times if incoming X was equal to 0
+	cpx	#$0
+	bne	docopy_loop
 	jsr	getput
 	dex
-	bne	docopy
-	dec	count+1
-	bne	docopy
+	bne	docopy_hack
+
+docopy_loop:
+	jsr	getput			; get/put byte
+	dex				; decrement count
+	bne	docopy_loop		; if not zero, loop
+docopy_hack:
+	dec	count+1			; if zero, decrement high byte
+	bne	docopy_loop		; if not zero, loop
+
 	rts
 
diff --git a/gr-sim/lz4d_verbose.c b/gr-sim/lz4d_verbose.c
index 0c51eb4a..248901de 100644
--- a/gr-sim/lz4d_verbose.c
+++ b/gr-sim/lz4d_verbose.c
@@ -39,22 +39,24 @@ void buildcount(void) {
 	ram[count+1]=x; 			// ??		// stx	count+1
 	cmp(0xf);				// if 15, more complicated // cmp	#$0f
 	if (z==0) goto done_buildcount;		// otherwise A is count // bne	++
-minus_buildcount:
+buildcount_loop:
 	ram[count]=a;				//-	sta	count
 //	printf("MBC ");
-	printf("\tADDITIONAL BUILDCOUNT ");
 	getsrc();				//jsr	getsrc
+	printf("\tADDITIONAL BUILDCOUNT 0x%x, adding 0x%x\n",a,ram[count]);
 	x=a;					//tax
 	c=0;					//clc
 	adc(ram[count]);			//adc	count
+	printf("\tGOT 0x%x c=%d\n",a,c);
 	if (c==0) goto skip_buildcount;		// bcc	+
 	ram[count+1]++;				//inc	count+1
 skip_buildcount:
-	x++;					//+	inx
-	if (x==0) goto minus_buildcount;	//beq	-
+	printf("\tUPDATED COUNT %02X%02X\n",ram[count+1],a);
+	x++;					// check if x is 255	//+	inx
+	if (x==0) goto buildcount_loop;		// if so, add in next byte //beq	-
 done_buildcount: ;				//++	rts
-	printf("\tBUILDCOUNT= r[c+1]=%02X r[c]=%02X a=%02X x=%02X\n",
-		ram[count+1],ram[count],a,x);
+	printf("\tBUILDCOUNT= %02X%02X r[c+1]=%02X r[c]=%02X a=%02X x=%02X\n",
+		ram[count+1],a,ram[count+1],ram[count],a,x);
 }
 
 
@@ -79,14 +81,26 @@ static void getput(void) {
 	putdst();				// ; fallthrough
 }
 
+// 202 -> 201 -> 100 -> 1FF
+// 201 -> 100 -> 1FF
+// 200 -> 2ff -> 2fe
+// 1ff -> 1fe
+
 static void docopy(void) {
+	printf("\tDOCOPY ENTRY: %02X%02X\n",ram[count+1],x);
 						// docopy:
+	if (x==0) {
+		getput();
+		x--;
+		goto dc_hack;
+	}
 
 docopy_label:
-	printf("\tDOCOPY %02X%02X: ",ram[count+1]-1,x);
+	printf("\tDOCOPY %02X%02X: ",ram[count+1],x);
 	getput();				// jsr	getput
 	x--;					// dex
 	if (x!=0) goto docopy_label;		// bne	docopy
+dc_hack:
 	ram[count+1]--;				// dec	count+1
 	if (ram[count+1]!=0) goto docopy_label;	//bne	docopy
 						//rts