From e51a120676ee49ecb2bce76eb9a8f08c0098ec9b Mon Sep 17 00:00:00 2001
From: Bernhard Urban <lewurm@gmail.com>
Date: Sun, 16 Jan 2011 15:37:08 +0100
Subject: [PATCH] progs/sum: 92 -> 74 cycles (thx @ loop unrollen)

---
 progs/sum.s | 42 ++++++++++++++++++++++++++++++++++++------
 1 file changed, 36 insertions(+), 6 deletions(-)

diff --git a/progs/sum.s b/progs/sum.s
index d762c60..c03df08 100644
--- a/progs/sum.s
+++ b/progs/sum.s
@@ -81,12 +81,42 @@ sum:
 	ldis r0, 0
 	addi r3, r1, 0
 	retzs- ; check len != 0
+
+	andx r1, 0x3
+
+	cmpi r1, 1
+	addideq r3, r3, 3
+	subideq r2, r2, 12
+	breq- loop1
+
+	cmpi r1, 2
+	addideq r3, r3, 2
+	subideq r2, r2, 8
+	breq- loop2
+
+	cmpi r1, 3
+	addideq r3, r3, 1
+	subideq r2, r2, 4
+	breq- loop3
+
+	addi r3, r3, 4
 loop:
-	;load arr data
+	subi r3, r3, 4
+	retzs-
 	ldw r4, 0(r2)
-	;sum += arr[i];
 	add r0, r0, r4
-	addi r2, r2, 4
-	subi r3, r3, 1
-	brnz+ loop
-	ret
+
+loop3:
+	ldw r4, 4(r2)
+	add r0, r0, r4
+
+loop2:
+	ldw r4, 8(r2)
+	add r0, r0, r4
+
+loop1:
+	ldw r4, 12(r2)
+	add r0, r0, r4
+
+	addi r2, r2, 4*4
+	br+ loop
-- 
2.25.1