RE: [PATCH 3.12 065/122] lib/checksum.c: fix carry in csum_tcpudp_nofold

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



From: Karl Beldan 
> On Tue, Feb 17, 2015 at 12:04:22PM +0000, David Laight wrote:
> > > +static inline u32 from64to32(u64 x)
> > > +{
> > > +	/* add up 32-bit and 32-bit for 32+c bit */
> > > +	x = (x & 0xffffffff) + (x >> 32);
> > > +	/* add up carry.. */
> > > +	x = (x & 0xffffffff) + (x >> 32);
> > > +	return (u32)x;
> > > +}
> >
> > As a matter of interest, does the compiler optimise away the
> > second (x & 0xffffffff) ?
> > The code could just be:
> > 	x = (x & 0xffffffff) + (x >> 32);
> > 	return x + (x >> 32);
> >
>
> On my side, from what I've seen so far, your version results in better
> assembly, esp. with clang, but my first version
> http://article.gmane.org/gmane.linux.kernel/1875407:
>         x += (x << 32) + (x >> 32);
> 	return (__force __wsum)(x >> 32);
> resulted in even better assembly, I just verified with gcc/clang,
> x86_64/ARM and -O1,2,3.

The latter looks to have a shorter dependency chain as well.
Although I'd definitely include a comment saying that it is equivalent
to the two lines in the current patch.

Does either compiler manage to use a rotate for the two shifts?
Using '(x << 32) | (x >> 32)' might convince it to do so.
That would reduce it to three 'real' instructions and a register rename.

	David

��.n��������+%������w��{.n�����������ܨ}���Ơz�j:+v�����w����ޙ��&�)ߡ�a����z�ޗ���ݢj��w�f





[Index of Archives]     [Linux Kernel]     [Kernel Development Newbies]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite Hiking]     [Linux Kernel]     [Linux SCSI]