[Bug target/68945] enable libcilkrts on SPARC
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68945 Bug 68945 depends on bug 70595, which changed state. Bug 70595 Summary: Cilk Plus testsuite needs massive cleanup https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70595 What|Removed |Added Status|ASSIGNED|RESOLVED Resolution|--- |WONTFIX
[Bug target/68945] enable libcilkrts on SPARC
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68945 Rainer Orth changed: What|Removed |Added Status|ASSIGNED|RESOLVED Resolution|--- |FIXED --- Comment #15 from Rainer Orth --- Fixed for 7.1.
[Bug target/68945] enable libcilkrts on SPARC
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68945 --- Comment #14 from Rainer Orth --- Author: ro Date: Thu Dec 1 14:48:49 2016 New Revision: 243112 URL: https://gcc.gnu.org/viewcvs?rev=243112=gcc=rev Log: Import libcilkrts Build 4467 (PR target/68945) PR target/68945 Merge from upstream, version 2.0.4467.0. Fix typo in git URL. * aclocal.m4, configure, Makefile.in: Regenerate. Added: trunk/libcilkrts/runtime/config/sparc/ trunk/libcilkrts/runtime/config/sparc/cilk-abi-vla.c trunk/libcilkrts/runtime/config/sparc/os-fence.h trunk/libcilkrts/runtime/config/sparc/os-unix-sysdep.c Modified: trunk/libcilkrts/ChangeLog trunk/libcilkrts/Makefile.am trunk/libcilkrts/Makefile.in trunk/libcilkrts/README trunk/libcilkrts/configure trunk/libcilkrts/configure.ac trunk/libcilkrts/configure.tgt trunk/libcilkrts/include/cilk/cilk_undocumented.h trunk/libcilkrts/include/internal/cilk_version.h trunk/libcilkrts/include/internal/rev.mk trunk/libcilkrts/runtime/cilk-abi.c trunk/libcilkrts/runtime/cilk_fiber-unix.cpp trunk/libcilkrts/runtime/cilk_fiber.h trunk/libcilkrts/runtime/config/arm/os-fence.h trunk/libcilkrts/runtime/except-gcc.cpp trunk/libcilkrts/runtime/global_state.cpp trunk/libcilkrts/runtime/global_state.h trunk/libcilkrts/runtime/jmpbuf.h trunk/libcilkrts/runtime/linux-symbols.ver trunk/libcilkrts/runtime/local_state.h trunk/libcilkrts/runtime/mac-symbols.txt trunk/libcilkrts/runtime/os-unix.c trunk/libcilkrts/runtime/record-replay.cpp trunk/libcilkrts/runtime/scheduler.c trunk/libcilkrts/runtime/scheduler.h trunk/libcilkrts/runtime/sysdep-unix.c
[Bug target/68945] enable libcilkrts on SPARC
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68945 --- Comment #13 from Rainer Orth --- I see my patch has been applied upstream. Thanks for that. It would be really great if it could be imported into the gcc tree for the GCC 7 release, either before stage 1 closes or, release managers permitting, even after that. Rainer
[Bug target/68945] enable libcilkrts on SPARC
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68945 --- Comment #12 from ro at CeBiTec dot Uni-Bielefeld.DE --- > --- Comment #11 from Eric Botcazou --- [...] >> * One thing I wonder about is runtime/config/sparc/os-fence.h: when using >> __sync_synchronize, gcc emits membar #StoreLoad, while Stefan's patch had >> membar#LoadLoad | #LoadStore | #StoreStore | #StoreLoad. It seems that >> all but #StoreLoad are no-ops for TSO SPARC CPUs, but I'd better get this >> right. > > __sync_synchronize emits the minimum memory barrier for the memory model, > which > is TSO on Solaris so only #StoreLoad is needed. The 4 flags are needed for > RMO > theoretically, but I'm not sure RMO ever existed in real life. Ok, so I'll keep the __sync_synchronize and inline asm versions in sync and submit the result upstream. Thanks. Rainer
[Bug target/68945] enable libcilkrts on SPARC
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68945 --- Comment #11 from Eric Botcazou --- > * In runtime/config/sparc/os-unix-sysdep.c (__cilkrts_getticks) I needed > different > 32- and 64-bit versions. I tested the result in standalone program which > just > printed the result. This looks good to me. > * One thing I wonder about is runtime/config/sparc/os-fence.h: when using > __sync_synchronize, gcc emits membar #StoreLoad, while Stefan's patch had > membar#LoadLoad | #LoadStore | #StoreStore | #StoreLoad. It seems that > all but #StoreLoad are no-ops for TSO SPARC CPUs, but I'd better get this > right. __sync_synchronize emits the minimum memory barrier for the memory model, which is TSO on Solaris so only #StoreLoad is needed. The 4 flags are needed for RMO theoretically, but I'm not sure RMO ever existed in real life.
[Bug target/68945] enable libcilkrts on SPARC
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68945 Rainer Orth changed: What|Removed |Added Attachment #38221|0 |1 is obsolete|| Attachment #38413|0 |1 is obsolete|| --- Comment #10 from Rainer Orth --- Created attachment 38451 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=38451=edit Revised patch Here's a version of the patch that is, I believe, ready for upstream submission. It's based on the upstream version as has just been merged into the gcc repo. Compared to the previous sparcv9 support, the assertions are gone to allow for an easier read. I've also replaced the times(3C) based implementation of __cilkrts_getticks by one reading %tick directly, which should work on all UltraSPARC CPUs, independent of O/S. I've replaced the _GNU_SOURCE guard around runtime/sysdep-unix.c (get_runtime_path) around the use dladdr by the result of a configure check. The previous version had been wrong (just defining _GNU_SOURCE) doesn't suddently materialize dladdr), and the function is present in Solaris libc. Perhaps Eric could have a look at the SPARC assembler code to double-check that I got it right: * In runtime/config/sparc/os-unix-sysdep.c (__cilkrts_getticks) I needed different 32- and 64-bit versions. I tested the result in standalone program which just printed the result. * One thing I wonder about is runtime/config/sparc/os-fence.h: when using __sync_synchronize, gcc emits membar #StoreLoad, while Stefan's patch had membar#LoadLoad | #LoadStore | #StoreStore | #StoreLoad. It seems that all but #StoreLoad are no-ops for TSO SPARC CPUs, but I'd better get this right. I also believe that only so little of Stefan's code is left (__cilkrts_short_pause mostly) to not require a separate assignement for upstream submission. If I'm mistaken, we'd better get that ball rolling. Rainer
[Bug target/68945] enable libcilkrts on SPARC
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68945 --- Comment #9 from Rainer Orth --- Created attachment 38413 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=38413=edit sparcv9 support patch This patch (on top of the previous one) fixes the sparcv9 failures reported before. As I'd suspected, the problem was that the code didn't take the SPARC V9 stack bias into account: in a couple of places, the stack and frame pointers in the __builtin_setjmp/__builtin_longjmp internal jmpbuf are overwritten with unbiased addresses. The patch includes both debugging code (assertions to check that the stack pointer is either biased or unbiased) and CILK_ADJUST_SP/CILK_UNADJUST_SP macros to turn a stack pointer from unbiased to biased form. The patch is enough to successfully run all Cilk Plus in the gcc and g++ testsuites successfully on sparc-sun-solaris2.10, but there may be code paths that also need adjustments. I'm just posting the patch (which is relative to upstream libcilkrts, which should be merged into mainline any day now) for reference and comments. Rainer
[Bug target/68945] enable libcilkrts on SPARC
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68945 --- Comment #8 from ro at CeBiTec dot Uni-Bielefeld.DE --- > --- Comment #7 from Stefan Teleman --- >> Since Stefan hasn't followed up and I'm currently looking at other >> libcilkrts issues anyway, I'm taking over from here. > > Had every intention to follow-up with a cleaned-up patch, but did not find > the time to re-do the patch. No worries: I'm familiar both with the gcc build system and the testsuite, so the changes fell out naturally. > I will look into the SIGBUS crashes on SPARC64 and see what i come up > with. Great: one of should be able to figure out what's going on. Rainer
[Bug target/68945] enable libcilkrts on SPARC
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68945 --- Comment #7 from Stefan Teleman --- > Since Stefan hasn't followed up and I'm currently looking at other > libcilkrts issues anyway, I'm taking over from here. Had every intention to follow-up with a cleaned-up patch, but did not find the time to re-do the patch. I will look into the SIGBUS crashes on SPARC64 and see what i come up with.
[Bug target/68945] enable libcilkrts on SPARC
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68945 Rainer Orth changed: What|Removed |Added Attachment #37050|0 |1 is obsolete|| --- Comment #6 from Rainer Orth --- Created attachment 38221 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=38221=edit Cleaned-up patch
[Bug target/68945] enable libcilkrts on SPARC
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68945 Rainer Orth changed: What|Removed |Added Status|NEW |ASSIGNED CC||ro at gcc dot gnu.org Depends on||70595 Assignee|unassigned at gcc dot gnu.org |ro at gcc dot gnu.org Target Milestone|--- |7.0 --- Comment #5 from Rainer Orth --- Since Stefan hasn't followed up and I'm currently looking at other libcilkrts issues anyway, I'm taking over from here. I think I've addressed all review comments. The necessary testsuite changes have been split out into PR testsuite/70595. I'm attaching my current patch, which allows all cilk-plus testcases to PASS on 32-bit Solaris 12/SPARC. The picture is different for 64-bit, though, where a couple of the execution tests FAIL in a way I cannot yet make sense of: they all die with SIGBUS like this: Thread 2 received signal SIGSEGV, Segmentation fault. [Switching to Thread 1 (LWP 1)] 0x7f614d14 in elf_rtbndr () from /usr/lib/sparcv9/ld.so.1 (gdb) where #0 0x7f614d14 in elf_rtbndr () from /usr/lib/sparcv9/ld.so.1 #1 0x7efd8e6c in _ti_bind_clear () from /lib/64/libc.so.1 #2 0x7f6335e8 in elf_bndr () from /usr/lib/sparcv9/ld.so.1 I suspect this is due to libcilkrts messing around with the internals of a __builtin_setjmp/__builtin_longjmp jmpbuf in libcilkplus/runtime/cilk_fiber-unix.cpp (cilk_fiber_sysdep::run), but haven't yet pinpointed what's actually wrong. Calls for more investigation. Rainer Referenced Bugs: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70595 [Bug 70595] Cilk Plus testsuite needs massive cleanup
[Bug target/68945] enable libcilkrts on SPARC
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68945 Jakub Jelinek changed: What|Removed |Added CC||jakub at gcc dot gnu.org --- Comment #4 from Jakub Jelinek --- For __EXTENSIONS*, you IMHO want to use autoconf AC_USE_SYSTEM_EXTENSIONS and remove the explicit _GNU_SOURCE defines in the sources. In os-unix-sysdep.c you are mixing up the defines, some stuff is SPARC related (e.g. the membar), but other stuff is Solaris related (the headers, getticks), you really should use different define for that, or different files depending on the size of changes. Otherwise it will fail to build on sparc*-linux*.
[Bug target/68945] enable libcilkrts on SPARC
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68945 Eric Botcazou changed: What|Removed |Added Status|UNCONFIRMED |NEW Last reconfirmed||2015-12-16 CC||ebotcazou at gcc dot gnu.org Summary|RFE: enable libcilkrts on |enable libcilkrts on SPARC |SPARC [ patch included ]| Ever confirmed|0 |1 Severity|normal |enhancement --- Comment #2 from Eric Botcazou --- Patches need to be posted on the gcc-patches@ mailing-list. A few remarks: - SPARC is not a platform/target, only an architecture, so I presume that you mean SPARC/Solaris here? - you don't need 2 entries in configure.tgt - you don't need 2 entries in the target field of testcases. - you cannot patch Makefile.in directly, go through configure.ac instead. - I don't think that patching runtime/config/generic/os-unix-sysdep.c is the way to go, you probably need to create runtime/config/sparc instead. - Why do you test all the possible flavours of macros? #if defined(sun) || defined(__sun) || defined(__sun__) #ifdef __sun__ is the correct version.
[Bug target/68945] enable libcilkrts on SPARC
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68945 --- Comment #3 from Stefan Teleman --- (In reply to Eric Botcazou from comment #2) > Patches need to be posted on the gcc-patches@ mailing-list. > > A few remarks: > > - SPARC is not a platform/target, only an architecture, so I presume that > you mean SPARC/Solaris here? Actually, I also mean to include Linux and the BSD's that still build on and publish for SPARC. I know that FreeBSD and Gentoo Linux still build and publish SPARC distros. > > - you don't need 2 entries in configure.tgt > > - you don't need 2 entries in the target field of testcases. > > - you cannot patch Makefile.in directly, go through configure.ac instead. > > - I don't think that patching runtime/config/generic/os-unix-sysdep.c is > the way to go, you probably need to create runtime/config/sparc instead. > > - Why do you test all the possible flavours of macros? > > #if defined(sun) || defined(__sun) || defined(__sun__) > > #ifdef __sun__ is the correct version. OK, thank you for your comments. I will correct these and resubmit to the gcc-patches@ mailing list.
