sh: Handle calling csum_partial with misaligned data

In rare circumstances csum_partial() can be called with data which is
not 16 or 32 bit aligned. This is been observed with RPC calls for NFS
file systems for example. Add support for handling this without resorting
to the misaligned fixup code (which is why this hasn't been seen as a
problem). This mimics the i386 version, which has had this support for
some time.
Signed-off-by: Stuart Menefy <stuart.menefy@st.com>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>

arch/sh/lib/checksum.S

@@ -36,8 +36,7 @@
  */
 
 /*	
- * unsigned int csum_partial(const unsigned char *buf, int len,
- *                           unsigned int sum);
+ * asmlinkage __wsum csum_partial(const void *buf, int len, __wsum sum);
  */
 
 .text
@@ -49,11 +48,31 @@ ENTRY(csum_partial)
 	   * Fortunately, it is easy to convert 2-byte alignment to 4-byte
 	   * alignment for the unrolled loop.
 	   */
-	mov	r5, r1
 	mov	r4, r0
-	tst	#2, r0		! Check alignment.
-	bt	2f		! Jump if alignment is ok.
+	tst	#3, r0		! Check alignment.
+	bt/s	2f		! Jump if alignment is ok.
+	 mov	r4, r7		! Keep a copy to check for alignment
 	!
+	tst	#1, r0		! Check alignment.
+	bt	21f		! Jump if alignment is boundary of 2bytes.
+
+	! buf is odd
+	tst	r5, r5
+	add	#-1, r5
+	bt	9f
+	mov.b	@r4+, r0
+	extu.b	r0, r0
+	addc	r0, r6		! t=0 from previous tst
+	mov	r6, r0
+	shll8	r6
+	shlr16	r0
+	shlr8	r0
+	or	r0, r6
+	mov	r4, r0
+	tst	#2, r0
+	bt	2f
+21:
+	! buf is 2 byte aligned (len could be 0)
 	add	#-2, r5		! Alignment uses up two bytes.
 	cmp/pz	r5		!
 	bt/s	1f		! Jump if we had at least two bytes.
@@ -61,16 +80,17 @@ ENTRY(csum_partial)
 	bra	6f
 	 add	#2, r5		! r5 was < 2.  Deal with it.
 1:
-	mov	r5, r1		! Save new len for later use.
 	mov.w	@r4+, r0
 	extu.w	r0, r0
 	addc	r0, r6
 	bf	2f
 	add	#1, r6
 2:
+	! buf is 4 byte aligned (len could be 0)
+	mov	r5, r1
 	mov	#-5, r0
-	shld	r0, r5
-	tst	r5, r5
+	shld	r0, r1
+	tst	r1, r1
 	bt/s	4f		! if it's =0, go to 4f
 	 clrt
 	.align	2
@@ -92,30 +112,31 @@ ENTRY(csum_partial)
 	addc	r0, r6
 	addc	r2, r6
 	movt	r0
-	dt	r5
+	dt	r1
 	bf/s	3b
 	 cmp/eq	#1, r0
-	! here, we know r5==0
-	addc	r5, r6			! add carry to r6
+	! here, we know r1==0
+	addc	r1, r6			! add carry to r6
 4:
-	mov	r1, r0
+	mov	r5, r0
 	and	#0x1c, r0
 	tst	r0, r0
-	bt/s	6f
-	 mov	r0, r5
-	shlr2	r5
+	bt	6f
+	! 4 bytes or more remaining
+	mov	r0, r1
+	shlr2	r1
 	mov	#0, r2
 5:
 	addc	r2, r6
 	mov.l	@r4+, r2
 	movt	r0
-	dt	r5
+	dt	r1
 	bf/s	5b
 	 cmp/eq	#1, r0
 	addc	r2, r6
-	addc	r5, r6		! r5==0 here, so it means add carry-bit
+	addc	r1, r6		! r1==0 here, so it means add carry-bit
 6:
-	mov	r1, r5
+	! 3 bytes or less remaining
 	mov	#3, r0
 	and	r0, r5
 	tst	r5, r5
@@ -139,8 +160,18 @@ ENTRY(csum_partial)
 8:
 	addc	r0, r6
 	mov	#0, r0
-	addc	r0, r6 
+	addc	r0, r6
 9:
+	! Check if the buffer was misaligned, if so realign sum
+	mov	r7, r0
+	tst	#1, r0
+	bt	10f
+	mov	r6, r0
+	shll8	r6
+	shlr16	r0
+	shlr8	r0
+	or	r0, r6
+10:
 	rts
 	 mov	r6, r0