I would say the switch to __sync atomics by default is surprising to me. I am *not* campaigning to reverse the result of the RFC (of which I mist have missed the details and outcome). However, I have a couple observations:
1) The loss of aqc/rel distinctions is probably minor on x86/x86-64 but might be significant on other platforms 2) For the case of x86 the "old" inline assembly provided opal_atomic_cmpset_64, and implemented 64-bit add, sub and swap in terms of it. However, gcc is NOT providing a __sync_bool_compare_and_swap_8(), or at least not where I tested (gcc 4.8.4 and 4.2.1, for instance). You could call that a gcc bug if you wish. -Paul On Mon, May 9, 2016 at 10:43 AM, Hjelm, Nathan Thomas <hje...@lanl.gov> wrote: > We have chosen to use the __sync builtins by default on master. There was > an rfc on it awhile ago. Is there a good reason to go back to the inline by > default or is this just surprising? > > > ________________________________________ > From: devel on behalf of Paul Hargrove > Sent: Monday, May 09, 2016 11:12:16 AM > To: Open MPI Developers > Subject: Re: [OMPI devel] Master broken for ILP32 > > Regarding "distro": > This was happening, for instance, on OpenBSD and NetBSD (32-bit kernels on > 64-bit h/w) when testing your PR1643. > However, it sounds like Nathan knows how/where to fix this. > > HOWEVER, that is not the only issue here. > Why is master is picking the BUILTIN (__sync) atomics (as shown in the > configure output quoted below), while v2.x (same system and same config > args) uses a .s file: > *** Assembler > checking dependency style of gcc -std=gnu99... gcc3 > checking for BSD- or MS-compatible name lister (nm)... /usr/bin/nm -B > checking the name lister (/usr/bin/nm -B) interface... BSD nm > checking for fgrep... /usr/bin/grep -F > checking if .proc/endp is needed... no > checking directive for setting text section... .text > checking directive for exporting symbols... .globl > checking for objdump... objdump > checking if .note.GNU-stack is needed... no > checking suffix for labels... : > checking prefix for global symbol labels... > checking prefix for lsym labels... .L > checking prefix for function in .type... @ > checking if .size is needed... yes > checking if .align directive takes logarithmic value... no > checking if processor supports x86_64 16-byte compare-and-exchange... no > checking if gcc -std=gnu99 supports GCC inline assembly... yes > checking if gcc -std=gnu99 supports DEC inline assembly... no > checking if gcc -std=gnu99 supports XLC inline assembly... no > checking for assembly format... default-.text-.globl-:--.L-@-1-0-1-1-0 > checking for assembly architecture... IA32 > checking for builtin atomics... BUILTIN_NO > checking for perl... perl > checking for pre-built assembly file... yes (atomic-ia32-linux-nongas.s) > checking for atomic assembly filename... atomic-ia32-linux-nongas.s > > > -Paul > > On Mon, May 9, 2016 at 1:22 AM, Gilles Gouaillardet <gil...@rist.or.jp > <mailto:gil...@rist.or.jp>> wrote: > > Paul, > > > on which distro are you running ? > > are you compiling on a 64 bit distro to generate a 32 bit library ? > > > it seems we are currently only testing a atomic on a long (32 bits on a 32 > bits arch) and > > then incorrectly assume it works also on 64 bits (!) > > > Cheers, > > > Gilles > > On 5/9/2016 3:59 PM, Paul Hargrove wrote: > Perhaps this is already known. > Several builds I've tried recently from the ompi (not ompi-release) repo > are failing on 32-bit platforms with > ../../../opal/.libs/libopen-pal.so: undefined reference to > `__sync_add_and_fetch_8' > > This is impacting PRs that I am being asked to test (e.g. 1643). > > Note that I did *not* configure with --enable-builtin-atomics, yet > configure seems to show them being selected anyway: > *** Assembler > checking dependency style of gcc -std=gnu99... gcc3 > checking for BSD- or MS-compatible name lister (nm)... /usr/bin/nm -B > checking the name lister (/usr/bin/nm -B) interface... BSD nm > checking for fgrep... /usr/bin/grep -F > checking for __sync builtin atomics... yes > checking for processor support of __sync builtin atomic compare-and-swap > on 128-bit values... no > checking for __sync builtin atomic compare-and-swap on 128-bit values with > -mcx16 flag... no > checking if .proc/endp is needed... no > checking directive for setting text section... .text > checking directive for exporting symbols... .globl > checking for objdump... objdump > checking if .note.GNU-stack is needed... no > checking suffix for labels... : > checking prefix for global symbol labels... > checking prefix for lsym labels... .L > checking prefix for function in .type... @ > checking if .size is needed... yes > checking if .align directive takes logarithmic value... no > checking if processor supports x86_64 16-byte compare-and-exchange... no > checking for assembly architecture... IA32 > checking for builtin atomics... BUILTIN_SYNC > checking for atomic assembly filename... none > > -Paul > > -- > Paul H. Hargrove <mailto:phhargr...@lbl.gov> > phhargr...@lbl.gov<mailto:phhargr...@lbl.gov> > Computer Languages & Systems Software (CLaSS) Group > Computer Science Department Tel: +1-510-495-2352 > <tel:%2B1-510-495-2352> > Lawrence Berkeley National Laboratory Fax: +1-510-486-6900 > <tel:%2B1-510-486-6900> > > > > _______________________________________________ > devel mailing list > de...@open-mpi.org<mailto:de...@open-mpi.org> > Subscription: https://www.open-mpi.org/mailman/listinfo.cgi/devel > Link to this post: > http://www.open-mpi.org/community/lists/devel/2016/05/18941.php > > > _______________________________________________ > devel mailing list > de...@open-mpi.org<mailto:de...@open-mpi.org> > Subscription: https://www.open-mpi.org/mailman/listinfo.cgi/devel > Link to this post: > http://www.open-mpi.org/community/lists/devel/2016/05/18942.php > > > > -- > Paul H. Hargrove phhargr...@lbl.gov<mailto: > phhargr...@lbl.gov> > Computer Languages & Systems Software (CLaSS) Group > Computer Science Department Tel: +1-510-495-2352 > Lawrence Berkeley National Laboratory Fax: +1-510-486-6900 > _______________________________________________ > devel mailing list > de...@open-mpi.org > Subscription: https://www.open-mpi.org/mailman/listinfo.cgi/devel > Link to this post: > http://www.open-mpi.org/community/lists/devel/2016/05/18946.php > -- Paul H. Hargrove phhargr...@lbl.gov Computer Languages & Systems Software (CLaSS) Group Computer Science Department Tel: +1-510-495-2352 Lawrence Berkeley National Laboratory Fax: +1-510-486-6900
