Improve speed of extended length loops

Author: calc84maniac

src/ce/lz4.src

@@ -127,18 +127,17 @@ lz4_decompress_block_internal:
 	ld	c, a
 	ld	b, 0
 	; Check for extended length
-	cp	a, $0F + 4
+	add	a, -($0F + 4)
 	jr	nz, .L.copy_backref
 	; Save the backref pointer and restore the input pointer
 	ex	(sp), hl
 .L.backref_loop:
-	; Read the next length byte
-	ld	a, (hl)
+	; Read the next length byte and add 1
+	adc	a, (hl)
 	inc	hl
 	; Decrement the length by 1
 	dec	bc
-	; Increment the length byte, and check if it was 255
-	inc	a
+	; Check if the length byte was 255
 	jr	nz, .L.copy_backref_long
 	; Increment the length by 256
 	inc	b
@@ -153,6 +152,8 @@ lz4_decompress_block_internal:
 	inc	bc
 	; Restore C
 	ld	c, a
+	; Set A = 0, preserving carry as 1
+	ld	a, b
 	jr	.L.backref_loop
 
 .L.literal_long:
@@ -161,14 +162,16 @@ lz4_decompress_block_internal:
 	; Start with literal length of 15
 	ld	b, 0
 	dec	bc
+.L.literal_loop_outer:
+	; Set A = 0, preserving carry as 1
+	ld	a, b
 .L.literal_loop:
-	; Read the next length byte
-	ld	a, (hl)
+	; Read the next length byte and add 1
+	adc	a, (hl)
 	inc	hl
 	; Decrement the length by 1
 	dec	bc
-	; Increment the length byte, and check if it was 255
-	inc	a
+	; Check if the length byte was 255
 	jr	nz, .L.copy_literal_long
 	; Increment the length by 256
 	inc	b
@@ -183,7 +186,7 @@ lz4_decompress_block_internal:
 	inc	bc
 	; Restore C
 	ld	c, a
-	jr	.L.literal_loop
+	jr	.L.literal_loop_outer
 
 .L.copy_literal_long:
 	; Copy the amount prior to the last length byte, minus 1