[v2,2/2] powerpc32: add support for csum_add()

[Scott Wood]

On Tue, 2015-04-28 at 21:01 +0200, christophe leroy wrote:
> 
> 
> Le 25/03/2015 02:30, Scott Wood a écrit :
> 
> > On Tue, Feb 03, 2015 at 12:39:27PM +0100, LEROY Christophe wrote:
> > > The C version of csum_add() as defined in include/net/checksum.h gives the
> > > following assembly:
> > >        0:       7c 04 1a 14     add     r0,r4,r3
> > >        4:       7c 64 00 10     subfc   r3,r4,r0
> > >        8:       7c 63 19 10     subfe   r3,r3,r3
> > >        c:       7c 63 00 50     subf    r3,r3,r0
> > > 
> > > include/net/checksum.h also offers the possibility to define an arch specific
> > > function.
> > > This patch provides a ppc32 specific csum_add() inline function.
> > What makes it 32-bit specific?
> > 
> > 
> As far as I understand, the 64-bit will do a 64 bit addition, so we
> will have to handle differently the carry, can't just be an addze like
> in 32-bit.

OK.  Before I couldn't find where this was ifdeffed to 32-bit, but it's
in patch 1/2.

> The generated code is most likely different on ppc64. I have no ppc64
> compiler so I can't check what gcc generates for the following code:
> 
> __wsum csum_add(__wsum csum, __wsum addend)
> {
> 	u32 res = (__force u32)csum;
> 	res += (__force u32)addend;
> 	return (__force __wsum)(res + (res < (__force u32)addend));
> }
> 
> Can someone with a ppc64 compiler tell what we get ?

With CONFIG_GENERIC_CPU:

   0xc000000000001af8 <+0>:	add     r3,r3,r4
   0xc000000000001afc <+4>:	cmplw   cr7,r3,r4
   0xc000000000001b00 <+8>:	mfcr    r4
   0xc000000000001b04 <+12>:	rlwinm  r4,r4,29,31,31
   0xc000000000001b08 <+16>:	add     r3,r4,r3
   0xc000000000001b0c <+20>:	clrldi  r3,r3,32
   0xc000000000001b10 <+24>:	blr

The mfcr is particularly nasty, at least on our chips.

With CONFIG_CPU_E6500:

   0xc000000000001b30 <+0>:	add     r3,r3,r4
   0xc000000000001b34 <+4>:	cmplw   cr7,r3,r4
   0xc000000000001b38 <+8>:	mfocrf  r4,1
   0xc000000000001b3c <+12>:	rlwinm  r4,r4,29,31,31
   0xc000000000001b40 <+16>:	add     r3,r4,r3
   0xc000000000001b44 <+20>:	clrldi  r3,r3,32
   0xc000000000001b48 <+24>:	blr

Ideal (short of a 64-bit __wsum) would probably be something like (untested):

	add	r3,r3,r4
	srdi	r5,r3,32
	add	r3,r3,r5
	clrldi	r3,r3,32

Or in C code (which would let the compiler schedule it better):

static inline __wsum csum_add(__wsum csum, __wsum addend)
{
        u64 res = (__force u64)csum;
        res += (__force u32)addend;
        return (__force __wsum)((u32)res + (res >> 32));
}

-Scott