Re: [PATCH] __div64_32: implement division by multiplication for 32-bit arches

["Måns Rullgård" <mans@mansr.com>]

Alexey Brodkin <Alexey.Brodkin@synopsys.com> writes:

> Hi mans,
>
> On Thu, 2015-10-29 at 12:52 +0000, Måns Rullgård wrote:
>> Alexey Brodkin <Alexey.Brodkin@synopsys.com> writes:
>> 
>> > Existing default implementation of __div64_32() for 32-bit arches unfolds
>> > into huge routine with tons of arithmetics like +, -, * and all of them
>> > in loops. That leads to obvious performance degradation if do_div() is
>> > frequently used.
>> > 
>> > Good example is extensive TCP/IP traffic.
>> > That's what I'm getting with perf out of iperf3:
>> >  -------------->8--------------
>> >     30.05%  iperf3   [kernel.kallsyms]        [k] copy_from_iter
>> >     11.77%  iperf3   [kernel.kallsyms]        [k] __div64_32
>> >      5.44%  iperf3   [kernel.kallsyms]        [k] memset
>> >      5.32%  iperf3   [kernel.kallsyms]        [k] stmmac_xmit
>> >      2.70%  iperf3   [kernel.kallsyms]        [k] skb_segment
>> >      2.56%  iperf3   [kernel.kallsyms]        [k] tcp_ack
>> >  -------------->8--------------
>> > 
>> > do_div() here is mostly used in skb_mstamp_get() to convert nanoseconds
>> > received from local_clock() to microseconds used in timestamp.
>> > BTW conversion itself is as simple as "/=1000".
>> > 
>> > Fortunately we already have much better __div64_32() for 32-bit ARM.
>> > There in case of division by constant preprocessor calculates so-called
>> > "magic number" which is later used in multiplications instead of divisions.
>> > It's really nice and very optimal but obviously works only for ARM
>> > because ARM assembly is involved.
>> > 
>> > Now why don't we extend the same approach to all other 32-bit arches
>> > with multiplication part implemented in pure C. With good compiler
>> > resulting assembly will be quite close to manually written assembly.
>> > 
>> > And that change implements that.
>> > 
>> > But there's at least 1 problem which I don't know how to solve.
>> > Preprocessor magic only happens if __div64_32() is inlined (that's
>> > obvious - preprocessor has to know if divider is constant or not).
>> > 
>> > But __div64_32() is already marked as weak function (which in its turn
>> > is required to allow some architectures to provide its own optimal
>> > implementations). I.e. addition of "inline" for __div64_32() is not an
>> > option.
>> > 
>> > So I do want to hear opinions on how to proceed with that patch.
>> > Indeed there's the simplest solution - use this implementation only in
>> > my architecture of preference (read ARC) but IMHO this change may
>> > benefit other architectures as well.
>> 
>> I tried something similar for MIPS a while ago after noticing a similar
>> perf report.  Adapting Nico's ARM code gave some nice speedups, but only
>> when I used MIPS assembly for the long multiplies.  Apparently gcc is
>> still too stupid to do the sane thing.
>
> Could you please elaborate a little bit on what was a problem with gcc
> compared to hand-written asm?

In the final multiplications (the ones using ARM assembly), gcc has a
tendency to multiply things by zero and add the (zero) result to
something.  This generally happens when multiplying a 64-bit value by a
32-bit one.  The 32-bit value is simply converted to 64-bit by the usual
promotion rules, and gcc forgets that the upper half is know to be zero.

> The point is if preprocessor does proper constant propagation then compiler
> will need to implement only calculations marked "run-time calculations".
> And in its turn those are pretty straight-forward 32-bit + and *.

The constant calculation is fine.  It's the final multiplication that's
the problem.

> And at least on ARC I saw with that change perf no longer captures
> __div64_32() during iperf and iperf results itself improved for about 10%.
> So I'd say advantage is quite noticeable.

There was an improvement without assembly as well, but with the MIPS
equivalent of the ARM assembly, it got much better.

-- 
Måns Rullgård
mans@mansr.com