Today's implementation of csum_shift() leads to branching based on
parity of 'offset'

	000002f8 <csum_block_add>:
	     2f8:	70 a5 00 01 	andi.   r5,r5,1
	     2fc:	41 a2 00 08 	beq     304 <csum_block_add+0xc>
	     300:	54 84 c0 3e 	rotlwi  r4,r4,24
	     304:	7c 63 20 14 	addc    r3,r3,r4
	     308:	7c 63 01 94 	addze   r3,r3
	     30c:	4e 80 00 20 	blr

Use first bit of 'offset' directly as input of the rotation instead of
branching.

	000002f8 <csum_block_add>:
	     2f8:	54 a5 1f 38 	rlwinm  r5,r5,3,28,28
	     2fc:	20 a5 00 20 	subfic  r5,r5,32
	     300:	5c 84 28 3e 	rotlw   r4,r4,r5
	     304:	7c 63 20 14 	addc    r3,r3,r4
	     308:	7c 63 01 94 	addze   r3,r3
	     30c:	4e 80 00 20 	blr

And change to left shift instead of right shift to skip one more
instruction. This has no impact on the final sum.

	000002f8 <csum_block_add>:
	     2f8:	54 a5 1f 38 	rlwinm  r5,r5,3,28,28
	     2fc:	5c 84 28 3e 	rotlw   r4,r4,r5
	     300:	7c 63 20 14 	addc    r3,r3,r4
	     304:	7c 63 01 94 	addze   r3,r3
	     308:	4e 80 00 20 	blr

Seems like only powerpc benefits from a branchless implementation.
Other main architectures like ARM or X86 get better code with
the generic implementation and its branch.
Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: David S. Miller <davem@davemloft.net>