Skip to contentSkip to conversation

common checksum code uses post-sh2 instruction

PATCH In review Rob Landley 2 participants 2 messages Sep 2, 2015 – Sep 4, 2015 SuperH

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 common checksum code uses post-sh2 instruction

From: Rob Landley <[email protected]>

The common checksum code uses shld which sh2 hasn't got, so unwind it.

Note: http://nommu.org/jcore added SHAD and SHLD, and if anybody's building for an actual sh2 they'll either need gcc 4.7 or earlier or a fix to https://gcc.gnu.org/bugzilla/show_bug.cgi?idT089 but when regression testing against the actual historical hardware, you need this.

---

 arch/sh/lib/checksum.S |   10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)
arch/sh/lib/checksum.S +6 -4
index 356c8ec..7f0fc6b 100644
@@ -88,8 +88,9 @@ ENTRY(csum_partial)
 2:
 	! buf is 4 byte aligned (len could be 0)
 	mov	r5, r1
-	mov	#-5, r0
-	shld	r0, r1
+	shlr2	r1		! mov	#-5, r0
+	shlr2	r1		! shld	r0, r1
+	shlr	r1
 	tst	r1, r1
 	bt/s	4f		! if it's =0, go to 4f
 	 clrt
@@ -288,8 +289,9 @@ DST(	mov.w	r0,@r5		)
 	addc	r0,r7
 2:
 	mov	r6,r2
-	mov	#-5,r0
-	shld	r0,r6
+	shlr2	r6		! mov	#-5, r0
+	shlr2	r6		! shld	r0, r6
+	shlr	r6
 	tst	r6,r6
 	bt/s	2f
 	 clrt
maintainer
#1 lore raw
Quoted 21 lines · Rob Landley

From: Rob Landley <[email protected]>

The common checksum code uses shld which sh2 hasn't got, so unwind it.

Note: http://nommu.org/jcore added SHAD and SHLD, and if anybody's building for an actual sh2 they'll either need gcc 4.7 or earlier or a fix to https://gcc.gnu.org/bugzilla/show_bug.cgi?idT089 but when regression testing against the actual historical hardware, you need this.

---

 arch/sh/lib/checksum.S |   10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)
arch/sh/lib/checksum.S +0 -0
index 356c8ec..7f0fc6b 100644

arch/sh/lib/checksum.S ENTRY(csum_partial) · line 93

Quoted 8 lines
arch/sh/lib/checksum.S +3 -2
@@ -88,8 +88,9 @@ ENTRY(csum_partial)
 2:
 	! buf is 4 byte aligned (len could be 0)
 	mov	r5, r1
-	mov	#-5, r0
-	shld	r0, r1
+	shlr2	r1		! mov	#-5, r0
+	shlr2	r1		! shld	r0, r1
+	shlr	r1

Is this a performance-critical code path (e.g. for networking)? If not, I like your current solution best because it's simple and #ifdef-free. But if it is critical and the shld is faster, perhaps #ifdef for SH3+ would be appropriate, and we could expand to include J2 in the #ifdef with the J2 support patches?

Rich