Re: communication pattern causes corrupted data --resolved

I thought I sent this out last week, but my mailer was acting flaky so
maybe it didn't go out.

Anyhow...its fixed!

Apparently 'configure --host=alphaev56....' was not enough to tell glibc
to compile for an ev56.  Looking at 'objdump -xd memcpy.o' showed lots
of ldq_u and stq_u instructions (load and store quad-word unaligned),
leading to the assumption that it was still compiled for the ev4.

After configuring ("--enable-add-ons --host=alphaev56-redhat-linux") and
recompiling glibc with the proper flags ("-mcpu=ev56 -freorder-blocks
-DNDEBUG=1 -g -O3") the problem went away.

Thanks go to Greg, David, and all the helpful folks on the list and at
Myricom.

-JE

On Tue, 2003-03-25 at 22:58, Greg Lindahl wrote:
> By the way, the quick way to see what memcopy() etc are actually using
> is to use ar to extract the appropriate .o from libc.a (you can figure
> out which .o using nm), and then run "objdump -xd foo.o" and read the
> assembly. Trust no one, keep your laser handy, and don't use the source.
> 
> -- greg
> 
> _______________________________________________
> Myrinet mailing list
> [email protected]
> http://email.osc.edu/mailman/listinfo/myrinet