harmony-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Rana Dasgupta" <rdasg...@gmail.com>
Subject Re: [drlvm] Another mfence operation??? fast_thread_array???
Date Sat, 21 Apr 2007 02:03:04 GMT
On 4/20/07, Nathan Beyer <ndbeyer@apache.org> wrote:
> I agree that we need to add the lock as well on windows. BTW - the
> locking for RW barriers was pulled from this document [1]. The lock
> instruction seemed the best of the three options; use mfence (not an
> option), use no-op lock or cpuid, which is apparently very slow.

Yes, I agree this is the safest if you really want a real read write
barrier that can be used anywhere, with least common denominator code
on x86, using a single binary image. But the penalty needs to be the
same on Windows and Linux.
>
> Another option would be looking at the operation just after the
> barrier; according to this [1], any lock-prefixed instruction will
> work on x86, so instead of a lock and wasted instruction, we could
> lock with the next instruction. Though, this may only be something a
> JIT would do.

Not all instructions can take the lock prefix that causes an assert on
the bus, which is what you need for the instruction to serialize, so
this "may" not always work.
>
> What does the EM64T mean?
> Isn't that just saying P4? I'm still not
> quite clear on the differences between EM6T, IPF, x86_64/x64, etc?

I agree, there are just too many acronyms that seem to be ever
changing!  EM64T here indicates 64 bit x86, also called AMD64
elsewhere, or x64. It does have a seperate binary for us, and for it
we use the VS2005 compiler. (For reference, your Linux part of the
patch seperates this out very nicely). Unfortunately, for VS2005, we
can't do __asm{ mfence } etc. since Microsoft does not support inline
assembly in this new compiler. So a seperate .asm file ( above ) is
needed. The build file will have to be modified to assemble this( see
other port/atomic/win etc. )directories, sorry I didn't add the build
change. So the above code will use mfence only on EM64T ( as in Linux
) , but use the "lock add..." on all 32 bit x86 platforms. IPF  ( also
referred to as iA64 ) is Itanium, and we only support it on Linux
where you use "asm volatile("mf:::memory" which is correct. We don't
need to worry about it on Windows.

>How
> do these map to the builds? There's not a separate EM64T end state,
> right?
>
> If that does just mean a P4, I don't think this code snippet is
> correct, unless we're going to start separating our builds that way.
>
> [1] http://gee.cs.oswego.edu/dl/jmm/cookbook.html
>
> -Nathan
>
> On 4/20/07, Rana Dasgupta <rdasgupt@gmail.com> wrote:
> > Nathan,
> >    I was going through your patch and had a small suggestion. We had
> > discussed on other related threads that on Windows we wanted to use
> > the MSVC provided _ReadWriteBarrier intrinsic and the patch does that,
> > but this may not be enough. This generates no instruction, but is a
> > compiler fence only. x86 does not need a store fence, but for smp the
> > load fence can be certainly needed.
> >   That's the assumption on Linux where in apr_memory_rw_barrier() we
> > insert both a compiler and an instruction fence, eg.,
> >
> > apr_memory_rw_barrier()
> > {
> > ...
> > #else
> > asm volatile ("lock; addl $0, 0(%%esp)":::"memory");
> > ...
> > }
> >
> > I agree that since there is no mfence on PIII, probably the dummy lock
> > is the safest option for x86 wide code. But  "lock" is expensive,
> > specially for Pentium 4 , Centrino etc. which have the "mfence"
> > instruction. And this penalty on current platforms is only on Linux.
> > My suggestion is to do the same on Windows as well. If we need the
> > real fence, we need it on both....So on Windows...
> >
> > On Windows in apr_thread_ext.c:
> > ================================
> > #if _MSC_VER < 1400
> >         extern void _ReadWriteBarrier();
> > #else
> >         #include <intrin.h>
> > #endif
> > #pragma intrinsic (_ReadWriteBarrier)
> >
> > APR_DECLARE(void) apr_memory_rw_barrier(){
> > #ifdef _EM64T_
> >         // create rwfence64() in asm file port\src\thread\win\rwfence.asm
> >         rwfence64();
> > #else
> >         __asm {lock add [esp], 0 }
> > #endif
> >        _ReadWriteBarrier();
> > }
> >
> > In port\src\thread\win\rwfence.asm:
> > ===================================
> > PUBLIC rwfence64
> >
> > _TEXT SEGMENT
> >
> > rwfence64 PROC
> >
> >         mfence
> >
> >
> > PROC ENDP
> > TEXT ENDS
> >
> > What do you think?
> >
> > Thanks,
> > Rana
> >
> > On 4/17/07, Nathan Beyer <ndbeyer@apache.org> wrote:
> > > I think I have this resolved with this commit [1]. I'm now able to run
> > > DRLVM in interpreted mode on my Quad P3 ... at least a 'Hello World'
> > > app.
> > >
> > > -Nathan
> > >
> > > [1] http://svn.apache.org/viewvc?view=rev&rev=529880
> > >
> > > On 4/18/07, Nathan Beyer <ndbeyer@apache.org> wrote:
> > > > I've run into another SSE2 operation in DRLVM, this time it's in the
> > > > hythr library, but I can't find the specific instance. According to
> > > > GDB, it's happening in fast_thread_array() in libhythr.so.
> > > >
> > > > Is this from the 'apr_thread_ext.c' file? I'm going to attempt to
> > > > modify this to be like the changes I made to the 'atomics.cpp', so
> > > > that this can run on P3 (SSE) CPUs.
> > > >
> > > > nathan@pants:~/harmony/drlvm-trunk/build/lnx_ia32_gcc_debug/deploy/jdk/jre/bin$
> > > > ./java -Xint -cp /home/nathan/workspace/helloworld/bin HelloWorld
> > > > Illegal instruction
> > > > nathan@pants:~/harmony/drlvm-trunk/build/lnx_ia32_gcc_debug/deploy/jdk/jre/bin$
> > > > gdb --args ./java -Xint -cp /home/nathan/workspace/helloworld/bin
> > > > HelloWorld
> > > > GNU gdb 6.6-debian
> > > > Copyright (C) 2006 Free Software Foundation, Inc.
> > > > GDB is free software, covered by the GNU General Public License, and you
are
> > > > welcome to change it and/or distribute copies of it under certain conditions.
> > > > Type "show copying" to see the conditions.
> > > > There is absolutely no warranty for GDB.  Type "show warranty" for details.
> > > > This GDB was configured as "i486-linux-gnu"...
> > > > Using host libthread_db library "/lib/tls/i686/cmov/libthread_db.so.1".
> > > > (gdb) run
> > > > Starting program:
> > > > /home/nathan/harmony/drlvm-trunk/build/lnx_ia32_gcc_debug/deploy/jdk/jre/bin/java
> > > > -Xint -cp /home/nathan/workspace/helloworld/bin HelloWorld
> > > > [Thread debugging using libthread_db enabled]
> > > > [New Thread -1214834000 (LWP 32414)]
> > > > [New Thread -1215992944 (LWP 32417)]
> > > > [New process 32414]
> > > > [New LWP 32414]
> > > >
> > > > Program received signal SIGILL, Illegal instruction.
> > > > 0xb7bc543f in fast_thread_array ()
> > > >    from /home/nathan/harmony/drlvm-trunk/build/lnx_ia32_gcc_debug/deploy/jdk/jre/bin/libhythr.so
> > > > (gdb) x/1i $eip
> > > > 0xb7bc543f <fast_thread_array+510367>:  mfence
> > > > (gdb)
> > > >
> > >
> >
>

Mime
View raw message