On x86-64, r238497 miscompiled 178.galgel in SPEC CPU 2000: [hjl@gnu-ivb-1 00000002]$ cat galgel.out.mis 0203: C1 = ( -0.3411 , 0.2388 ) C1 = ( -0.3475 , 0.2414 ) ^ 0268: Parameter Mu = 0.1048E+05 Parameter Mu = 0.1068E+05 ^ 0269: Parameter Tau = -0.1024E-01 Parameter Tau = -0.1041E-01 ^ 0270: Floquet exponent = -0.6821 Floquet exponent = -0.6950 ^ [hjl@gnu-ivb-1 00000002]$
Reduced test case?
The miscompare of 178.galgel is observed in aarch64-linux as well.
bifg21.f90 is miscompiled.
If there is a bug, it is likely an array dependence that is mishandled.
Also occurring on powerpc64.
Is anybody working on reducing this? AFAIK, none of the gfortran developers has access to SPEC.
(In reply to Thomas Koenig from comment #6) > Is anybody working on reducing this? > > AFAIK, none of the gfortran developers has access to SPEC. Reducing it would probably prove hard as the bug is a miscompile and therefore only shows at runtime in a >30 second test. H.J. has narrowed the miscompilation down to one file, would it be helpful for you if I generated some diffs of when the dumps first change, and the generated assembly for x86_64?
also miscompiles CP2K, but haven't been able to narrow it down.
(In reply to James Greenhalgh from comment #7) > Reducing it would probably prove hard as the bug is a miscompile and > therefore only shows at runtime in a >30 second test. H.J. has narrowed the > miscompilation down to one file, would it be helpful for you if I generated > some diffs of when the dumps first change, and the generated assembly for > x86_64? Use -Warray-temporaries, using a compiler with and one without the patch. The compiler with the patch should be missing a warning about creating an array temporary. Create a minimum test case containing just the offending line, plus any declarations that are needed to make it valid. Check if the two compilers still exhibit the difference with -Warray-temporaries. Voila, minimum test case! BTW, I will only have time to work on this in around two weeks. With a test case, it would be OK with me if somebody reverted the patch. I can then rework it to take care of that particular bug.
(In reply to Thomas Koenig from comment #9) > With a test case, it would be OK with me if somebody reverted the > patch. I can then rework it to take care of that particular bug. A revert would be good I think.. this is a small testcase showing the wrong results and the missing warning. I suspect it could be matmul specific. > cat test.f90 REAL, DIMENSION(:,:), POINTER :: a REAL, DIMENSION(:,:), ALLOCATABLE :: b ALLOCATE(a(4,4),b(4,2)) CALL RANDOM_NUMBER(a) CALL RANDOM_NUMBER(b) a(1:4,1:2)=MATMUL(a(1:4,1:4),b(1:4,1:2)) WRITE(6,*) a(1,1) END > gfortran -O0 -Warray-temporaries test.f90 ; ./a.out test.f90:6:11: a(1:4,1:2)=MATMUL(a(1:4,1:4),b(1:4,1:2)) 1 Warning: Creating array temporary at (1) [-Warray-temporaries] 0.770401359 > gfortran -O1 -Warray-temporaries test.f90 ; ./a.out 0.515214324
This even gives wrong results at -O0 ... > cat test.f90 INTEGER, DIMENSION(:,:), POINTER :: a INTEGER, DIMENSION(:,:), ALLOCATABLE :: b ALLOCATE(a(4,4),b(4,2)) a=1 ; b=2 a(:,1:2)=MATMUL(a(:,1:4),b(:,:)) write(6,*) a IF (ANY(a.NE.RESHAPE((/8,8,8,8,8,8,8,8,1,1,1,1,1,1,1,1/),(/4,4/)))) & CALL ABORT END gives correct results with gcc 5.3
Just for completeness sake, this is what I whittled bifg21.f90 down to. r238496 produces a warning, r238497 does not. I'll also note 178.galgel also failed at -O0, on powerpc64 at least. Subroutine BifG21 (G21, W11, W20, V, U, NS) Real*8, Dimension(1000,1000) :: HtTim Integer N, K Real*8, Allocatable, Dimension(:) :: POP2 Np = Max0(N,K) Allocate( POP2(Np) ) POP2(1:N) = MATMUL( HtTim(1:K,1:K), POP2(1:K) ) Deallocate( POP2 ) Return End > gfortran -S -m64 junk.f90 -Warray-temporaries junk.f90:10:23: POP2(1:N) = MATMUL( HtTim(1:K,1:K), POP2(1:K) ) 1 Warning: Creating array temporary at (1) [-Warray-temporaries]
This looks OK. A revert is pre-approved (if that is even needed), so whowever can do so at the moment, feel free. Just make sure to repoen PR 71902.
Author: renlin Date: Thu Jul 28 11:21:53 2016 New Revision: 238815 URL: https://gcc.gnu.org/viewcvs?rev=238815&root=gcc&view=rev Log: [PATCH] Revert Revert r238497 because of PR 71961. This patch reverts the change for PR 71902 since it causes 178.gagel miscompile in spec2000 as reported in PR 71961 which was observed in x86_64, aarch64, powerpc64. gcc/fortran/ChangeLog: 2016-07-28 Renlin Li <renlin.li@arm.com> Revert 2016-07-19 Thomas Koenig <tkoenig@gcc.gnu.org> PR fortran/71902 * dependency.c (gfc_check_dependency): Use dep_ref. Handle case if identical is true and two array element references differ. (gfc_dep_resovler): Move most of the code to dep_ref. (dep_ref): New function. * frontend-passes.c (realloc_string_callback): Name temporary variable "realloc_string". gcc/testsuite/ChangeLog: 2016-07-28 Renlin Li <renlin.li@arm.com> Revert 2016-07-19 Thomas Koenig <tkoenig@gcc.gnu.org> PR fortran/71902 * gfortran.dg/dependency_47.f90: New test. Removed: trunk/gcc/testsuite/gfortran.dg/dependency_47.f90 Modified: trunk/gcc/fortran/ChangeLog trunk/gcc/fortran/dependency.c trunk/gcc/fortran/frontend-passes.c trunk/gcc/testsuite/ChangeLog
The change r238497 has been reverted as r238815. I confirmed that, after the revert, the 178.gagel mis-compare is fixed in aarch64-linux environment. PR 71902 is reopend as well.
Author: tkoenig Date: Sun Aug 7 13:24:32 2016 New Revision: 239220 URL: https://gcc.gnu.org/viewcvs?rev=239220&root=gcc&view=rev Log: 2016-08-07 Thomas Koenig <tkoenig@gcc.gnu.org> PR fortran/71961 * gfortran.dg/matmul_10.f90: New testcase. Added: trunk/gcc/testsuite/gfortran.dg/matmul_10.f90 Modified: trunk/gcc/testsuite/ChangeLog
Test case committed, closing.
Author: tkoenig Date: Sun Aug 7 15:25:56 2016 New Revision: 239221 URL: https://gcc.gnu.org/viewcvs?rev=239221&root=gcc&view=rev Log: 2016-08-07 Thomas Koenig <tkoenig@gcc.gnu.org> PR fortran/71961 * gfortran.dg/pr70040.f90: New testcase. Added: trunk/gcc/testsuite/gfortran.dg/pr70040.f90 Modified: trunk/gcc/testsuite/ChangeLog