Re: [PATCH] riscv: fix memmove and optimise memcpy when misalign

[Bin Meng]

On Tue, May 25, 2021 at 10:37 PM Gary Guo <gary@xxxxxxxxxxx> wrote:
>
> On Sun, 23 May 2021 17:12:23 +0000
> David Laight <David.Laight@xxxxxxxxxx> wrote:
>
> > From: Palmer Dabbelt
> > > Sent: 23 May 2021 02:47
> > ...
> > > IMO the right way to go here is to just move to C-based string
> > > routines, at least until we get to the point where we're seriously
> > > optimizing for specific processors.  We went with the C-based
> > > string rountines in glibc as part of the upstreaming process and
> > > found only some small performance differences when compared to the
> > > hand-written assembly, and they're way easier to maintain.
>
> I prefer C versions as well, and actually before commit 04091d6 we are
> indeed using the generic C version. The issue is that 04091d6
> introduces an assembly version that's very broken. It does not offer
> and performance improvement to the C version, and breaks all processors
> without hardware misalignment support (yes, firmware is expected to
> trap and handle these, but they are painfully slow).
>
> I noticed the issue because I ran Linux on my own firmware and found
> that kernel couldn't boot. I didn't implement misalignment emulation at
> that time (and just send the trap to the supervisor).

Similarly, I noticed exactly the same issue as what you saw on
mainline U-Boot SPL that it suddenly does not boot on SiFive
Unleashed. Bisect leads to the commit that brought the kernel buggy
assembly mem* version into U-Boot.

> Because 04091d6 is accepted, my assumption is that we need an assembly
> version.

That's my understanding as well. I thought kernel folks prefer
hand-written assembly over pure C :)

> So I spent some time writing, testing and optimising the assembly.

Thank you for the fix!

>
> > >
> > > IIRC Linux only has trivial C string routines in lib, I think the
> > > best way to go about that would be to higher performance versions
> > > in there. That will allow other ports to use them.
> >
> > I certainly wonder how much benefit these massively unrolled
> > loops have on modern superscaler processors - especially those
> > with any form of 'out of order' execution.
> >
> > It is often easy to write assembler where all the loop
> > control instructions happen in parallel with the memory
> > accesses - which cannot be avoided.
> > Loop unrolling is so 1970s.
> >
> > Sometimes you need to unroll once.
> > And maybe interleave the loads and stores.
> > But after that you can just be trashing the i-cache.
>
> I didn't introduce the loop unrolling though. The loop unrolled
> assembly is there before this patch, and I didn't even change the
> unroll factor. I only added a path to handle misaligned case.
>
> There are a lot of diffs because I did made some changes to the
> register allocation so that the code is more optimal. I also made a few
> cleanups and added a few comments. It might be easier to review if you
> apply the patch locally and just look at the file.

I would vote to apply this patch, unless we wanted to do "git revert
04091d6", but that would break KASAN I guess?

Regards,
Bin