harmony-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Alexey Varlamov" <alexey.v.varla...@gmail.com>
Subject Re: problem under windows
Date Fri, 25 May 2007 04:15:06 GMT
Small success story is here:
I found an AthlonXP+WinXP box and tried debug build with /arch:SSE.
Jitrino loaded successfully, but vmi.dll and both GCs appeared to
depend on MSVCR71D.DLL which is not in the harmony bundle and does not
exist in system w/o MSVC installed. After replacing those dlls with
release versions, I was able to work on Eclipse and observed no
errors.
Though it was painfully slow I must admit :(. I hope HARMONY-3737 will
improve client experience, + we may consider downgrading jitted code
for non-SSE2-capable CPUs to SSE at least, not to x87 level.

Next I tried to run Forte3.0CE installer and met assertion in
Ia32i586InstsExpansion.cpp - success story ended ;).

I'll create JIRAs for the met problems.
--
Alexey


2007/5/25, Nathan Beyer <ndbeyer@apache.org>:
> BTW - I do have P3 machines for testing. One's running W2K and the
> other's running Ubuntu.
>
> -Nathan
>
> On 5/24/07, Elford, Chris L <chris.l.elford@intel.com> wrote:
> > FYI, I have created http://issues.apache.org/jira/browse/HARMONY-3961
> > To reflect this desired behavior.
> >
> > Regards,
> > Chris Elford
> > Intel SSG/Enterprise Solutions Software Division
> >
> > -----Original Message-----
> > From: Elford, Chris L [mailto:chris.l.elford@intel.com]
> > Sent: Thursday, May 24, 2007 9:46 AM
> > To: dev@harmony.apache.org
> > Subject: RE: problem under windows
> >
> > Interesting.  Its good to hear that some thought has been given to P3
> > class systems.  I would think /arch:SSE would be okay.  This would let
> > us run on both P3s and Athlon (like the one that Michael was trying to
> > run on) which support both MMX and SSE but not SSE2.
> >
> > Is there a compiler option that generates back-off MMX code on platforms
> > that don't support SSE or SSE2 instructions?  If so, it would result in
> > larger executables and might impact code locality but it might be a good
> > tradeoff.
> >
> > Regards,
> >
> > Chris Elford
> > Intel SSG/Enterprise Solutions Software Division
> >
> > -----Original Message-----
> > From: Mikhail Fursov [mailto:mike.fursov@gmail.com]
> > Sent: Thursday, May 24, 2007 2:12 AM
> > To: dev@harmony.apache.org
> > Subject: Re: problem under windows
> >
> > Well, it looks like we need to remove this line from jitrino.dll build
> > configuration for 32bit platform.
> >                 <compilerarg value="/arch:SSE2" />
> >
> > On 5/24/07, Mikhail Fursov <mike.fursov@gmail.com> wrote:
> > >
> > > On 5/24/07, Elford, Chris L <chris.l.elford@intel.com> wrote:
> > > >
> > > > Hi all,
> > > >
> > > >   I tracked down a system that has SSE but not SSE2 (a dual
> > processor P3
> > > > system).  For me, adding "-xe ii" to the ntsd command line let the
> > > > debugger break in at the first chance exception.  It reports that
> > what I
> > > >
> > > > believe is an SSE2 instruction does indeed happen during jitrino.dll
> > > > startup.
> > > >
> > > > 01365fb0 f20f100520553901 movsd xmm0,qword ptr
> > > > [jitrino!JIT_gc_start+0xd5810 (01395520)]
> > > >
> > > > Without jitrino symbol information ( e.g., a pdb file for
> > jitrino.dll),
> > > > its hard to tell what function is being executed (it is unlikely
> > that it
> > > > is JIT_gc_start.  It would be nice if the snapshot builds included
> > > > symbol files to help make for more meaningful stack traces.
> > > >
> > > > It may be good to discuss [again?] on the lists whether more people
> > can
> > > > come into the project to help create legacy modes that earlier
> > > > instruction sets (and earlier Oses) can still execute without
> > crippling
> > > > current instruction set processors (and/or OSes).
> > > >
> > > > I don't see an open JIRA about this one.  The closest I see is
> > > > http://issues.apache.org/jira/browse/HARMONY-3246 which is more
> > about
> > > > code that the JIT creates rather than the JIT itself.
> > > >
> > > > Ntsd log below showing more detail around faulting instruction.
> > > > Mikhail?
> > >
> > >
> > > You're right that we fixed Jitrino.OPT to generate SSE independent
> > code
> > > (use x87) if SSE2 is not available.
> > > Another JIRA (not commited yet):
> > http://issues.apache.org/jira/browse/HARMONY-3737
> > > makes JET to compile all methods without FPU insts (>90% of methods)
> > and
> > > pass other methods to OPT
> > > I checked dumpbin /disasm for jitrino.dll and see a lot of SSE/SSE2
> > insts
> > > in code.
> > >
> > > It's possible that our compiler options for jitrino.dll are too
> > > aggressive. I will check it and report.
> > >
> > > --
> > > Mikhail Fursov
> >
> >
> >
> >
> > --
> > Mikhail Fursov
> >
>

Mime
View raw message