Module Name: src
Committed By: matt
Date: Tue Dec 18 14:08:25 UTC 2012
Modified Files:
src/sys/arch/arm/arm: cpu_in_cksum_v4hdr.S
Log Message:
Add ldrd support (one additional instruction over 5 ldrs) but two of them
won't be executed resulting in 2 less loads which should save a few cyles.
To generate a diff of this commit:
cvs rdiff -u -r1.1 -r1.2 src/sys/arch/arm/arm/cpu_in_cksum_v4hdr.S
Please note that diffs are not public domain; they are subject to the
copyright notices on the relevant files.
Modified files:
Index: src/sys/arch/arm/arm/cpu_in_cksum_v4hdr.S
diff -u src/sys/arch/arm/arm/cpu_in_cksum_v4hdr.S:1.1 src/sys/arch/arm/arm/cpu_in_cksum_v4hdr.S:1.2
--- src/sys/arch/arm/arm/cpu_in_cksum_v4hdr.S:1.1 Tue Dec 18 13:41:42 2012
+++ src/sys/arch/arm/arm/cpu_in_cksum_v4hdr.S Tue Dec 18 14:08:25 2012
@@ -29,18 +29,30 @@
#include <machine/asm.h>
-RCSID("$NetBSD: cpu_in_cksum_v4hdr.S,v 1.1 2012/12/18 13:41:42 matt Exp $")
+RCSID("$NetBSD: cpu_in_cksum_v4hdr.S,v 1.2 2012/12/18 14:08:25 matt Exp $")
ENTRY(cpu_in_cksum_v4hdr)
+#ifdef _ARM_ARCH_DWORD_OK
+ tst r0, #4 /* 64-bit aligned? */
+ ldreqd r2, [r0], #8 /* load 1st/2nd words */
+ ldrne ip, [r0], #4 /* load 1st word */
+ ldreq ip, [r0, #8] /* load 5th word */
+ ldrned r2, [r0, #8] /* load 4th/5th words */
+#else
ldr ip, [r0] /* load 1st word */
ldr r3, [r0, #4] /* load 2nd word */
- adds r3, r3, ip /* add 1st to 2nd */
ldr r2, [r0, #8] /* load 3rd word */
- adcs r2, r2, r3 /* add sum to 3rd */
+#endif
+ adds r3, r3, ip /* accumulate */
+ adcs r2, r2, r3 /* accumulate */
+#ifdef _ARM_ARCH_DWORD_OK
+ ldrd r0, [r0] /* load remaining words */
+#else
ldr r1, [r0, #12] /* load 4th word */
- adcs r1, r1, r2 /* add sum to 4th */
ldr r0, [r0, #16] /* load 5th word */
- adcs r0, r0, r1 /* add sum to 5th */
+#endif
+ adcs r1, r1, r2 /* accumulate */
+ adcs r0, r0, r1 /* accumulate */
/*
* We now have a 33-bit (r0 + carry) sum which needs to resolved to a
* 16-bit sum.
@@ -62,7 +74,7 @@ ENTRY(cpu_in_cksum_v4hdr)
movw r1, #0xffff /* load 0xffff */
#else
mov r1, #0x10000 /* load 0x10000 */
- sub r1, r1, #1 /* subtract 1 to get 0xffff */
+ sub r1, r1, #1 /* subtract by 1 to get 0xffff */
#endif
subge r0, r0, r1 /* subtract 0xffff */
eor r0, r0, r1 /* complement lower halfword */
