[PATCH 1/1] riscv: prevent pipeline stall in __asm_to/copy_from_user

[Akira Tsukamoto]

On Sat, Jun 12, 2021 at 1:05 PM Palmer Dabbelt <palmer at dabbelt.com> wrote:
>
> On Tue, 08 Jun 2021 04:31:40 PDT (-0700), David.Laight at ACULAB.COM wrote:
> > From: Akira Tsukamoto
> >> Sent: 04 June 2021 10:57
> >>
> >> Reducing pipeline stall of read after write (RAW).
> >>
> >> These are the results from combination of the speedup with
> >> Gary's misalign fix. Speeds up from 680Mbps to 900Mbps.
> >>
> >> Before applying these two patches.
> >
> > I think the changes should be in separate patches.
> > Otherwise it is difficult to see what is relevant.
> > It also looks as if there is a register rename.
> > Maybe that should be a precursor patch?
>
> Yes, and I'd also prefer the original patches.  This also doesn't apply.
>
> > ...
> >
> > I think this is the old main copy loop:
> >>  1:
> >> -    fixup REG_L, t2, (a1), 10f
> >> -    fixup REG_S, t2, (a0), 10f
> >> -    addi a1, a1, SZREG
> >> -    addi a0, a0, SZREG
> >> -    bltu a1, t1, 1b
> > and this is the new one:
> >>  3:
> >> +    fixup REG_L a4,       0(a1), 10f
> >> +    fixup REG_L a5,   SZREG(a1), 10f
> >> +    fixup REG_L a6, 2*SZREG(a1), 10f
> >> +    fixup REG_L a7, 3*SZREG(a1), 10f
> >> +    fixup REG_L t0, 4*SZREG(a1), 10f
> >> +    fixup REG_L t1, 5*SZREG(a1), 10f
> >> +    fixup REG_L t2, 6*SZREG(a1), 10f
> >> +    fixup REG_L t3, 7*SZREG(a1), 10f
> >> +    fixup REG_S a4,       0(t5), 10f
> >> +    fixup REG_S a5,   SZREG(t5), 10f
> >> +    fixup REG_S a6, 2*SZREG(t5), 10f
> >> +    fixup REG_S a7, 3*SZREG(t5), 10f
> >> +    fixup REG_S t0, 4*SZREG(t5), 10f
> >> +    fixup REG_S t1, 5*SZREG(t5), 10f
> >> +    fixup REG_S t2, 6*SZREG(t5), 10f
> >> +    fixup REG_S t3, 7*SZREG(t5), 10f
> >> +    addi a1, a1, 8*SZREG
> >> +    addi t5, t5, 8*SZREG
> >> +    bltu a1, a3, 3b
> >
> > I don't know the architecture, but unless there is a stunning
> > pipeline delay for memory reads a simple interleaved copy
> > may be fast enough.
> > So something like:
> >       a = src[0];
> >       do {
> >               b = src[1];
> >               src += 2;
> >               dst[0] = a;
> >               dst += 2;
> >               a = src[0];
> >               dst[-1] = b;
> >       } while (src != src_end);
> >       dst[0] = a;
> >
> > It is probably worth doing benchmarks of the copy loop
> > in userspace.
>
> I also don't know this microarchitecture, but this seems like a pretty
> wacky load-use delay.
>
> Can we split out the misaligned handling fix to get that in sooner,
> that's likely the more urgent issue.

Sure, just give me a few days.

Akira

>
> >
> >       David
> >
> > -
> > Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
> > Registration No: 1397386 (Wales)