Re: [PATCH 1/2] riscv: Add memmove string operation.

[Palmer Dabbelt]

On Tue, 13 Aug 2019 19:22:15 PDT (-0700), Paul Walmsley wrote:
> On Tue, 13 Aug 2019, Palmer Dabbelt wrote:
>
> > On Mon, 12 Aug 2019 08:04:46 PDT (-0700), Christoph Hellwig wrote:
> > > On Wed, Aug 07, 2019 at 03:19:14PM +0800, Nick Hu wrote:
> > > > There are some features which need this string operation for compilation,
> > > > like KASAN. So the purpose of this porting is for the features like KASAN
> > > > which cannot be compiled without it.
> > > >
> > > > KASAN's string operations would replace the original string operations and
> > > > call for the architecture defined string operations. Since we don't have
> > > > this in current kernel, this patch provides the implementation.
> > > >
> > > > This porting refers to the 'arch/nds32/lib/memmove.S'.
> > >
> > > This looks sensible to me, although my stringop asm is rather rusty,
> > > so just an ack and not a real review-by:
> > >
> > > Acked-by: Christoph Hellwig <hch@xxxxxx>
> >
> > FWIW, we just write this in C everywhere else and rely on the compiler to
> > unroll the loops.  I always prefer C to assembly when possible, so I'd prefer
> > if we just adopt the string code from newlib.  We have a RISC-V-specific
> > memcpy in there, but just use the generic memmove.
> >
> > Maybe the best bet here would be to adopt the newlib memcpy/memmove as generic
> > Linux functions?  They're both in C so they should be fine, and they both look
> > faster than what's in lib/string.c.  Then everyone would benefit and we don't
> > need this tricky RISC-V assembly.  Also, from the look of it the newlib code
> > is faster because the inner loop is unrolled.
>
> There's a generic memmove implementation in the kernel already:
>
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/include/linux/string.h#n362

That ends up at __builtin_memcpy(), which ends up looking for memcpy() for 
large copies, which is in lib/string.c.  The code in there is just byte at a 
time memcpy()/memmove(), which is way slower than the newlib stuff.

> Nick, could you tell us more about why the generic memmove() isn't
> suitable?
>
>
> - Paul