[Vm-dev] Stack VMs with primitives for AndreasSystemProfiler?

Fri Mar 1 00:07:25 UTC 2013

On Wed, Feb 27, 2013 at 6:12 PM, Juan Vuletich <juanlists at jvuletich.org>wrote:

>
> Hi Eliot,
>
> >  On Wed, Feb 27, 2013 at 10:03 AM, Eliot Miranda
> > <eliot.miranda at gmail.com>wrote:
> >
> >> Hi Juan,
> >>
> >> On Wed, Feb 27, 2013 at 8:45 AM, Juan Vuletich (mail lists) <
> >> juanlists at jvuletich.org> wrote:
> >>
> >>>
> >>> Quoting "David T. Lewis" <lewis at mail.msen.com>:
> >>>
> >>>
> >>>> On Wed, Feb 27, 2013 at 08:19:20AM -0300, Juan Vuletich (mail lists)
> >>>> wrote:
> >>>>
> >>>>>
> >>>>> Hi Folks,
> >>>>>
> >>>>> In Cuis 4.1 (
> >>>>> https://github.com/jvuletich/**Cuis/blob/master/**
> >>>>> Cuis4WithLatestUpdates/Cuis4.**1-1619.zip?raw=true<
> https://github.com/jvuletich/Cuis/blob/master/Cuis4WithLatestUpdates/Cuis4.1-1619.zip?raw=true
> >)
> >>>>> I included AndreasSystemProfiler. But sometimes it is good to use a
> >>>>> StackVM for profiling, to avoid Cog inlining and get results that are
> >>>>> easier to
> >>>>> follow.
> >>>>>
> >>>>> The Mac and Windows VMs referenced from squeakVM.org don't include
> >>>>> the
> >>>>> required primitives. The only VMs with those primitives I found are
> >>>>> CogVMs from Eliot's site.
> >>>>>
> >>>>>
> >>>> Thanks,
> >>>>
> >>>> I just added a mantis issue for this:
> >>>>
> >>>> http://bugs.squeak.org/view.**php?id=7746<
> http://bugs.squeak.org/view.php?id=7746>
> >>>>
> >>>> Eliot, can you give me a quick pointer to the primitives in Cog that
> >>>> need
> >>>> to be added to the interpreter VM? Thanks.
> >>>>
> >>>>  Does anybody know if there are plans to add those primitives to
> >>>>> StackVMs?
> >>>>>
> >>>>>
> >>>> A terminology note, this issue is for the interpreter VM. The stack VM
> >>>> is
> >>>> part of Cog, and I expect that a stack VM will already contain the
> >>>> necessary
> >>>> primitives (though I did not check).
> >>>>
> >>>
> >>> Thanks for the correction. I understand that the interpreter VMs will
> >>> have the primitives in the future. Thank you!
> >>>
> >>> Let me restate, then. Does anybody know if there are recent StackVMs
> >>> precompiled for major platforms? I could not find any at Eliot愀 site.
> >>>
> >>
> >> I could be encouraged to build and add these to my site.  But I wonder
> >> why
> >> you need them.  What's wrong with the Cog VMs?
> >>
> >
> > First of all, Cog doesn't do inlining; it only does good inline cacheing,
> > including memoing whether a send is an MNU.  Second, one can disable the
> > JIT using -cogmaxlits -1 on the command line.  You'll get something
> slower
> > than the StackVM (because it'll be asking the jit whether methods should
> > be
> > jit all the time) but you should get values close to the StackVM.
>
> Thanks for the correction wrt inlining. There's nothing wrong with Cog.
> Cog is wonderful. Also thanks for the -cogmaxlits -1 tip. I'm sure I'll
> find use for it.
>
> Let me explain my problem.
>
> Tim's recent feedback on performance in Raspberry PI made me want to go do
> some profiling on a slow machine. The slowest machine I have at hand is a
> Thinkpad T23, Pentium III, 1.1GHz. I use a program called "Cpukiller3",
> that can slow it down to (for example) Raspi performance levels. Besides,
> I integrated AndreasSystemProfiler in Cuis, and I prefer using it. Now I
> need a suitable Windows VM:
> - CogVM: Can't run on the P-III, as it doesn't include SSE2 instructions.
> - StackVM: Not currently built by you. I don't know if it would require
> SSE2, anyway.
>

I certainly intended that neither of these would need SSE2, but that they
would use SSE2 if available.  SSE2 instructions are used in the Stack and
Cog VMs for locking/atomic instructions used in things like signalling
external semaphores.  The code in platforms/Cross/vm/sqAtomicOps.h is
supposed to fall back on non-SSE2 instructions if SSE2 is not defined (but
I've not tested it recently).  So try building on Windows with SSE2 not
defined and you may get lucky.

In Cog, SSE2 is further used to implement fast floating-point primitives.
 If the processor doesn't have SSE2 the Cog doesn't generate the
primitives.  So the thing to try is building the current sources but
deleting -msse2 from the line in cygwinbuild/Makefile:

CFLAGS:= $(INCLUDEPATH) -msse2 -ggdb2 -mwindows -mthreads -mno-cygwin
-mwin32 \
    -mno-rtd -mms-bitfields -mno-accumulate-outgoing-args $(OFLAGS)
$(NOBUILTIN)

If you try this let me know either way.  I can add the description to
HowToBuild.

> - InterpreterVM a.k.a. ClosureVM. Currently doesn't include the new
> profiling primitives. Will include them, eventually. Right now, this is
> what I'm using, limited to less precise MessageTally. It is also what I've
> been using for several years to look for code to optimize in Cuis.
>
> Additionally, wrt Cog, on 2013/1/23 you wrote:
> "That said there are still limitations with primitives and Cog.  Currently
> Cog only samples "interpreter" primitives.  Those primitives it implements
> in machine code (integer and float arithmetic, closure evaluation, at:,
> identityHash) are not sampled and won't show up; they will be charged to
> the calling method.".
> I'm not sure if this is a real issue. And I don't know if this limitation
> applies to StackVM or InterpreterVM. Can you please clarify?
>
> Another option would be some software for that can slow down an Atom or
> Core processor in a reliable and predictable way, and just use Cog. So
> far, I couldn't find any. Suggestions welcome!
>
> Finally, I could take the dust off my XO (300MHz Geode processor) and do
> profiling there. It is already slow enough, and can run Cog.
>
> Thanks in advance, Eliot.
>
> >
> >>
> >>>
> >>>
> >>>  Dave
> >>>>
> >>>>  Thanks,
> >>>>> Juan Vuletich
> >>>>>
> >>>>
> >>>>
> >>>
> >>>
> >>> Cheers,
> >>> Juan Vuletich
> >>>
> >>>
> >>
> >>
> >> --
> >> best,
> >> Eliot
> >>
> >
> >
> >
> > --
> > best,
> > Eliot
> >
>
> Cheers,
> Juan Vuletich
>

-- 
best,
Eliot
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.squeakfoundation.org/pipermail/vm-dev/attachments/20130228/ff5d3fe6/attachment.htm