Re: [AMBER-Developers] In which serial Amber 11 is compiled on a Mac running Snow Leopard

From: Ben Roberts <>
Date: Thu, 01 Oct 2009 18:45:42 -0400

case wrote:
> On Thu, Oct 01, 2009, Ben Roberts wrote:
>>> The Intel compilers [give a compiler error on amber.c]
> Please (generally) post the output of "icc --version" when reporting
> compiler errors. Also, when saying "on a Mac", indicate the OS version.
> So many things changed between 10.5 and 10.6 that it is hard to keep things
> straight.
> This is certainly weird in the extreme. This code hasn't changed in decades,
> and every other version of icc has been able to compile it with the default
> optimzation level. :-(
> On the other hand, we could certainly put -O0 into CFLAGS for intel compilers.
> It was the intent that only OCFLAGS stuff would be optimized, but if the intel
> default is -O2, then effectively CFLAGS and OCFLAGS give the same behavior.
> So, downgrading CFLAGS to -O0 should be OK.
> Of course, there is no reason for anyone to need to compile leap with intel
> (instead of gnu) compilers. But the compiler *will* make a difference for
> nab, so I guess it is good to test all this.

Right you are. Yep, I'm running Snow Leopard (10.6.1), and icc 11.1.067

Because it's a compiler internal error, I asked a real computer expert,
who tells me it's properly an Intel bug, not a LeAP one. I'm in the
process of figuring out how to raise it with the Intel developers, but
perhaps that would be better done by someone who has the same OS and
Intel-compiler version and is an actual LeAP developer (if such a person
exists and is willing to take this on board). If not, I can do it,
though trying to figure out which part of amber.c the compiler has a fit
over could prove interesting.

I checked the man pages to make sure, and -O2 is the Intel default these
days, unless any of various other compiler flags are used (like -fast,
which forces -O3 among other things, or various options which force
-O0). But I don't think we make a habit of using those flags.

On the positive side, this seems to be the only part of the code that
causes an optimisation-related glitch under Intel, and in general the
tests don't seem to fall over. (Some tests do fail, but I think they're
mostly the same that fail under the GNU compilers; I'll put up some
Bugzillas about them.) So I guess at least we don't absolutely need to
bust our collective boilers to turn LeAP back into a gcc-built program.


AMBER-Developers mailing list
Received on Thu Oct 01 2009 - 16:00:03 PDT
Custom Search