Wise decision. "Platform heals" went out with the '80's. Ken
-----Original Message----- From: devel-boun...@open-mpi.org [mailto:devel-boun...@open-mpi.org] On Behalf Of Jeff Squyres Sent: Tuesday, July 27, 2010 5:40 PM To: Open MPI Developers Subject: Re: [OMPI devel] [OMPI users] Debug info on Darwin In good email tradition, this thread went downhill. :-) So Brian and I got on the phone and had a good long discussion about it. The short version is that we need to ask those who know about compilers and debuggers to figure out what the Right Thing to do is. I'm not going to hold up 1.5.0 for this fix (i.e., revert out what was put in 1.5; I already reverted it out of the trunk). There is a (sucky but workable) workaround for Totalview/DDT support on OSX: ./configure CFLAGS=-gstabs+ ... We can get the real fix in 1.5.1 after getting some advice from mo' smarter people. On Jul 27, 2010, at 6:29 PM, Jeff Squyres (jsquyres) wrote: > Actually, it occurs to me that my commit doesn't even fix the case that it was intended to fix -- @#$%@#$%%@#$ !!! > > The main point of the patch was to fix totalview support. I did this by examining CFLAGS. But because CFLAGS doesn't always contain -g, we use $DEBUGGER_CFLAGS (setup in orte/config/orte_setup_debugger_flags.m4) to always compile the "special" .c files in libompi that *must* be compiled with -g (and friends) for totalview/ddt to work. > > My patch didn't address DEBUGGER_CFLAGS at all -- it only addressed CFLAGS, so it doesn't guarantee that the ompi/debugger/*.c files are compiled with -gstabs+. Sigh. So regardless, my first commit is busted. > > Here's a summary of random points: > > - Per above, we clearly need to add -gstabs+ for the DEBUGGER_CFLAGS -- not CFLAGS. But it *needs* to be added -- even if it's just for those ompi/debugger/*.c files. > ===> This is the most important point, I think. > ===> Is there a way to test to see if -gstabs+ is broken? > > - I agree that configure should always "just work" out of the box. We shouldn't add something that may be broken in some environment. > > - I didn't think we cared about OS X 10.3 anymore (OMPI's readme says that the earliest OS X we have tested is 10.4). Brian -- are you saying that you have a special environment that uses OS X 10.3 that we just broke? > > - FWIW: it is useful for OMPI *developers* to have -gstabs+. I'd venture that most end users don't care about -gstabs+ -- if their debuggers just step over MPI function calls (because they can't step in), they don't care. > > > > On Jul 27, 2010, at 6:15 PM, Ralph Castain wrote: > > > I do sympathize with the "the user said to do it" argument as that is in keeping with our approach elsewhere. IIRC, what Jeff had implemented does print out a warning if we override the flag, so this would only be a minor change that might help alert people to what is going on. I would also suggest an FAQ entry as well. > > > > > > On Jul 27, 2010, at 4:08 PM, Barrett, Brian W wrote: > > > > > Ok, so there is some middle ground I would dislike but not hate enough to object to. How about having a AC_MSG_WARN if we find -g in the CFLAGS/CXXFLAGS on OS X version 10.4 or later? Users get educated, I can still make executables that are debugabble for my weird environment, and this thread can die. It's not perfect, but at least it seems to accomplish everyone's goals. > > > > > > Brian > > > > > > On Jul 27, 2010, at 4:02 PM, Barrett, Brian W wrote: > > > > > >> Unfortunately, there really is no middle ground. THe only option is to ask Apple to fix -g to mean -gstabs on OS X. I'm cross compiling from one version to another, so running an executable won't work. Looking at the three or four ways that you can specify a target version won't work (especially since at least in 10.4, you could change them without command line or environment variables). > > >> > > >> Sorry, there's no middle ground - this patch broke a use case that used to work. I know you guys didn't know about -gstabs. That doesn't make it right to do evil things. > > >> > > >> Brian > > >> > > >> On Jul 27, 2010, at 3:47 PM, Ralph Castain wrote: > > >> > > >>> Can I offer a middle ground? I believe we have been burned enough with the -g vs -gstabs situation on OSX that it merits defaulting "appropriately". So... > > >>> > > >>> Can we detect if gstabs is actually supported by the OS vs the compiler? > > >>> > > >>> If not, can we add logic that checks the OS target version and "does the right thing"? > > >>> > > >>> My concern here is that our users are no more informed than Jeff or I were, and will almost certainly specify -g when what they really mean is "I want a debuggable executable". Unfortunately, as we all know, that isn't what you get on OSX once you hit 10.4 and above. > > >>> > > >>> > > >>> > > >>> On Jul 27, 2010, at 3:29 PM, Barrett, Brian W wrote: > > >>> > > >>>> On Jul 27, 2010, at 3:13 PM, Jeff Squyres wrote: > > >>>> > > >>>>> On Jul 27, 2010, at 5:02 PM, Barrett, Brian W wrote: > > >>>>> > > >>>>>> Yes, but say I'm using a custom built version of gcc that doesn't do -gstabs quite right. Now you've screwed me. > > >>>>> > > >>>>> The configure test checks to see if -gstabs+ is accepted by the compiler. > > >>>> > > >>>> Yes, but acceptance and usefulness are not the same thing. > > >>>> > > >>>> If the TARGET_VERSION is 10.3, the compiler will accept -gstabs, but the executable won't be debugable on 10.3, because 10.3 didn't use gstabs. So I really, really want -g and you've just prevented me from doing what I want. That's not just bad, it's awful. > > >>>> > > >>>>>> I'm a firm believer that our configure script should do what the user says, as exactly as possible. Changing AC behavior a little bit is a gray area, but one I'm almost ok with (since AC_PROG_CC will add -g if CFLAGS is empty). > > >>>>> > > >>>>> We override that, though -- we take out that -g if CFLAGS was empty. > > >>>>> > > >>>>> I understand what you're saying, and in general I agree -- that we should add as little as possible to what the user said. But I don't quite know how to balance: > > >>>>> > > >>>>> * adding as few flags as possible > > >>>>> * making debugging symbols work for those of us who aren't > > >>>>> familiar enough with OSX to know that you need the > > >>>>> special/secret -gstabs+ flag (and just expect -g to be enough) > > >>>>> > > >>>>> I've been developing POSIX software on a Mac for several years (i.e., not Mac-specific software, so I never dived into the details of Mac-specific functionality) and fell into the 2nd category until about a week ago. > > >>>>> > > >>>>> Got any suggestions? > > >>>> > > >>>> Yes, don't cause problems - if the user specified -g, assume he meant -g. I understand what you're saying, but I'm against changing what the user specified because we know better. We usually don't, and in this case, there are known use cases where that's true. > > >>>> > > >>>> Brian > > >>>> > > >>>> -- > > >>>> Brian W. Barrett > > >>>> Dept. 1423: Scalable System Software Sandia National > > >>>> Laboratories > > >>>> > > >>>> > > >>>> > > >>>> > > >>>> > > >>>> _______________________________________________ > > >>>> devel mailing list > > >>>> de...@open-mpi.org > > >>>> http://www.open-mpi.org/mailman/listinfo.cgi/devel > > >>> > > >>> > > >>> _______________________________________________ > > >>> devel mailing list > > >>> de...@open-mpi.org > > >>> http://www.open-mpi.org/mailman/listinfo.cgi/devel > > >>> > > >> > > >> -- > > >> Brian W. Barrett > > >> Dept. 1423: Scalable System Software Sandia National Laboratories > > >> > > >> > > >> > > > > > > -- > > > Brian W. Barrett > > > Dept. 1423: Scalable System Software > > > Sandia National Laboratories > > > > > > > > > > > > > > > > > > _______________________________________________ > > > devel mailing list > > > de...@open-mpi.org > > > http://www.open-mpi.org/mailman/listinfo.cgi/devel > > > > > > _______________________________________________ > > devel mailing list > > de...@open-mpi.org > > http://www.open-mpi.org/mailman/listinfo.cgi/devel > > > > > -- > Jeff Squyres > jsquy...@cisco.com > For corporate legal information go to: > http://www.cisco.com/web/about/doing_business/legal/cri/ > > > _______________________________________________ > devel mailing list > de...@open-mpi.org > http://www.open-mpi.org/mailman/listinfo.cgi/devel > -- Jeff Squyres jsquy...@cisco.com For corporate legal information go to: http://www.cisco.com/web/about/doing_business/legal/cri/ _______________________________________________ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel
