Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > comp.lang.java.programmer > #3300

Re: Java left shift and right shift operators.

Show key headers only | View raw

On 4/26/2011 10:40 PM, Travers Naran wrote:
...
> ;; if (b > 0) return a << b else return a >> -b
>  compare b to 0
>  jump-less-than to L1
>  a := a << b
>  jump L2
> L1:
>  b := -b
>  a := a >> b
> L2:
>  return  ;; result in a
...
> It's the jump that's adding the time. But on an x86, that kind of
> instruction can be entirely cached and the Intel/AMD processors can
> perform predictive branching to speed up this bit of code.
>
> The only way you should be getting 20-50 times longer is if you
> aren't using any sort of optimizing run-time. Example: running always
> as byte code without compiling.
>
> But usually this kind of if-checking and bit-twiddling is so fast
> that it's not worth optimizing beyond a single statement. What are
> you coding that makes this the most expensive calculation in your
> program?
...

The obvious possibility is a random mix of left and right shifts, with
the two directions equally probable. In that case, the hardware branch
prediction would mis-predict, on average, half the time.

When the hardware branch prediction gets it wrong, the result is a
complete pipeline flush. With deep pipelines and multiple instructions
in each pipeline stage, a pipeline flush costs the equivalent of dozens
of simple arithmetic operations.

My question is whether the measurement was from the actual program, or
from a benchmark that may differ from the real program in ways that
affect branch prediction.

> If branching is "20-50 times longer" than any of the operations
> (negation or shifting), how could you re-write this assembly to use a
> single branch statement?

Why bother? Your sample code only has one conditional branch,
"jump-less-than to L1". Unconditional jumps to a fixed target, such as
"jump L2" are always correctly predicted, and take about the same time
as a simple arithmetic operation. The method is short enough that the
JVM will in-line it anywhere it is frequently called so we don't have to
worry about the return.

Patricia

Back to comp.lang.java.programmer | Previous | Next — Previous in thread | Next in thread | Find similar

Thread

Java left shift and right shift operators. Sanny <softtanks22@hotmail.com> - 2011-04-26 00:38 -0700
  Re: Java left shift and right shift operators. Lothar Kimmeringer <news200709@kimmeringer.de> - 2011-04-26 09:55 +0200
    Re: Java left shift and right shift operators. Patricia Shanahan <pats@acm.org> - 2011-04-26 03:16 -0700
    Re: Java left shift and right shift operators. Lew <noone@lewscanon.com> - 2011-04-26 09:49 -0400
      Re: Java left shift and right shift operators. Travers Naran <tnaran@gmail.com> - 2011-04-26 07:09 -0700
        Re: Java left shift and right shift operators. Lew <noone@lewscanon.com> - 2011-04-26 11:21 -0400
        Re: Java left shift and right shift operators. Martin Gregorie <martin@address-in-sig.invalid> - 2011-04-26 15:37 +0000
      Re: Java left shift and right shift operators. Sanny <softtanks22@hotmail.com> - 2011-04-26 09:32 -0700
        Re: Java left shift and right shift operators. Patricia Shanahan <pats@acm.org> - 2011-04-26 11:09 -0700
          Re: Java left shift and right shift operators. Patricia Shanahan <pats@acm.org> - 2011-04-26 13:09 -0700
        Re: Java left shift and right shift operators. Travers Naran <tnaran@gmail.com> - 2011-04-26 22:40 -0700
          Re: Java left shift and right shift operators. Patricia Shanahan <pats@acm.org> - 2011-04-27 05:34 -0700
            Re: Java left shift and right shift operators. Owen Jacobson <angrybaldguy@gmail.com> - 2011-04-28 00:37 -0400
              Re: Java left shift and right shift operators. Patricia Shanahan <pats@acm.org> - 2011-04-28 16:02 -0700
              Re: Java left shift and right shift operators. Michael Wojcik <mwojcik@newsguy.com> - 2011-04-28 19:22 -0400
          Re: Java left shift and right shift operators. bugbear <bugbear@trim_papermule.co.uk_trim> - 2011-04-27 14:28 +0100
          Re: Java left shift and right shift operators. Thomas Richter <thor@math.tu-berlin.de> - 2011-04-30 00:02 +0200
      Re: Java left shift and right shift operators. Roedy Green <see_website@mindprod.com.invalid> - 2011-04-27 19:25 -0700
        Re: Java left shift and right shift operators. Lew <noone@lewscanon.com> - 2011-04-27 23:08 -0400
  Re: Java left shift and right shift operators. Sanny <softtanks22@hotmail.com> - 2011-04-26 02:04 -0700
    Re: Java left shift and right shift operators. Eric Sosman <esosman@ieee-dot-org.invalid> - 2011-04-26 07:34 -0400
      Re: Java left shift and right shift operators. Lew <noone@lewscanon.com> - 2011-04-26 09:47 -0400
        Re: Java left shift and right shift operators. Patricia Shanahan <pats@acm.org> - 2011-04-27 09:23 -0700
          Re: Java left shift and right shift operators. Lew <noone@lewscanon.com> - 2011-04-27 14:49 -0400
            Re: Java left shift and right shift operators. Paul Cager <paul.cager@googlemail.com> - 2011-04-27 17:28 -0700
              Re: Java left shift and right shift operators. Lew <noone@lewscanon.com> - 2011-04-27 22:06 -0400
      Re: Java left shift and right shift operators. Sanny <softtanks22@hotmail.com> - 2011-04-26 09:39 -0700
        Re: Java left shift and right shift operators. Patricia Shanahan <pats@acm.org> - 2011-04-26 11:17 -0700
        Re: Java left shift and right shift operators. Lew <noone@lewscanon.com> - 2011-04-26 14:55 -0400
          Re: Java left shift and right shift operators. Lew <noone@lewscanon.com> - 2011-04-26 14:57 -0400
          Re: Java left shift and right shift operators. Daniele Futtorovic <da.futt.news@laposte-dot-net.invalid> - 2011-04-26 21:38 +0200
          Re: Java left shift and right shift operators. Andreas Leitgeb <avl@gamma.logic.tuwien.ac.at> - 2011-04-26 23:32 +0000
            Re: Java left shift and right shift operators. Lew <noone@lewscanon.com> - 2011-04-27 00:08 -0400
              Re: Java left shift and right shift operators. Andreas Leitgeb <avl@gamma.logic.tuwien.ac.at> - 2011-04-27 06:57 +0000
                Re: Java left shift and right shift operators. Arved Sandstrom <asandstrom3minus1@eastlink.ca> - 2011-04-27 07:00 -0300
                Re: Java left shift and right shift operators. Lew <noone@lewscanon.com> - 2011-04-27 08:42 -0400
                Re: Java left shift and right shift operators. Andreas Leitgeb <avl@gamma.logic.tuwien.ac.at> - 2011-04-27 14:03 +0000
                Re: Java left shift and right shift operators. Lew <noone@lewscanon.com> - 2011-04-27 14:54 -0400
        Re: Java left shift and right shift operators. Eric Sosman <esosman@ieee-dot-org.invalid> - 2011-04-26 21:14 -0400
    Re: Java left shift and right shift operators. Travers Naran <tnaran@gmail.com> - 2011-04-26 07:30 -0700
      Re: Java left shift and right shift operators. Lew <noone@lewscanon.com> - 2011-04-26 11:24 -0400
        Re: Java left shift and right shift operators. Lew <noone@lewscanon.com> - 2011-04-26 11:35 -0400
          Re: Java left shift and right shift operators. Sanny <softtanks22@hotmail.com> - 2011-04-26 09:46 -0700
            Re: Java left shift and right shift operators. Lew <noone@lewscanon.com> - 2011-04-26 15:02 -0400
      Re: Java left shift and right shift operators. Lothar Kimmeringer <news200709@kimmeringer.de> - 2011-04-28 09:14 +0200
  Re: Java left shift and right shift operators. Joshua Cranmer <Pidgeot18@verizon.invalid> - 2011-04-26 08:41 -0400
  Re: Java left shift and right shift operators. Robert Klemme <shortcutter@googlemail.com> - 2011-04-26 22:06 +0200

csiph-web