Description: This test case exhibits the problem that the presence of a SECTIONS directive with only one SECTION inside of a PARALLEL region causes only one thread to be created when omp_set_num_threads was previously called with 4 threads. If the directives internal to this PARALLEL region are commented out, 4 threads are created and the program produces expected output. The compiler used was GNU gfortran. The test case works as expected when compiled with PGI. > gfortran -v Using built-in specs. Target: x86_64-suse-linux Configured with: ../xt-gcc-4.2.0/configure --prefix=/opt/gcc/4.2.0/snos --disable-nls --libdir=/opt/gcc/4.2.0/snos/lib --enable-languages=c,c++,fortran --with-gxx-include-dir=/opt/gcc/4.2.0/snos/include/g++ --with-slibdir=/opt/gcc/4.2.0/snos/lib --with-system-zlib --enable-shared --enable-__cxa_atexit x86_64-suse-linux Thread model: posix gcc version 4.2.0 20070514 (rpm:4) > cat bug2831.f90 ! Derived from OpenMP test omp1/F2_1_1_2_1c.f90 use omp_lib implicit none integer, parameter :: NT = 4 integer :: nth !$ call omp_set_dynamic(.false.) !$ call omp_set_num_threads(NT) !$omp parallel default(none) shared(nth) print *, omp_get_thread_num(), omp_get_num_threads() !$omp sections !$omp section nth=omp_get_num_threads() !$omp endsections !$omp endparallel print *, 'nth=',nth,' NT=',NT END > diff bug2831.f90 bug2831a.f90 11,12d10 < !$omp sections < !$omp section 14d11 < !$omp endsections Incorrect output is produced: > ftn -O0 -fopenmp -o x bug2831.f90 /opt/xt-pe/2.1/bin/snos64/ftn: INFO: linux target is being used > aprun -n 1 ./x 0 1 nth= 1 NT= 4 Application 217361 resources: utime 0, stime 0 Correct output from program with sections/section directives removed: > ftn -O0 -fopenmp -o xa bug2831a.f90 /opt/xt-pe/2.1/bin/snos64/ftn: INFO: linux target is being used > aprun -n 1 ./xa 0 4 1 4 3 4 2 4 nth= 4 NT= 4 Application 217362 resources: utime 0, stime 0 ------ Note: ftn is an alias for: /opt/gcc/4.2.0/bin/../snos/bin/gfortran -static -v -I/opt/xt-mpt/2.1/mpich2-64/GP/include -I/opt/xt-mpt/2.1/mpich2-64/GP/include -L/opt/xt-mpt/2.1/mpich2-64/GP/lib -I/opt/acml/3.6.1/gnu64/include -I/opt/xt-libsci/10.1.0/gnu/snos64/include -I/opt/xt-libsci/10.1.0/gnu/snos64/include/superlu -I/opt/xt-mpt/2.1/sma/P/include -L/opt/acml/3.6.1/gnu64/lib -L/opt/xt-libsci/10.1.0/gnu/snos64/lib -L/opt/xt-mpt/2.1/sma/P/lib -lmpichf90 -lsci -lacml -lsma -lmpichf90 -lmpich -lrt -D__CRAYXT_COMPUTE_LINUX_TARGET -D__TARGET_LINUX__ -fno-second-underscore -I/notbackedup/users/rsrel/rs64.DEV.070604.Mon/install/include -I/opt/xt-catamount/2.1/catamount/linux/include -I/opt/xt-service/2.1/include -L/notbackedup/users/rsrel/rs64.DEV.070604.Mon/install/lib/snos64 -L/opt/xt-pe/2.1/cnos/linux/64/lib -L/opt/xt-mpt/2.1/lib/snos64 -L/opt/xt-service/2.1/lib/snos64 -Wl,--start -lpct -lalpslli -lalpsutil -lportals -lpthread -Wl,--end -lgfortranbegin -lgfortran -lm
Toying with the example shows that the number of threads in the parallel region always equals the minimum of the number of SECTIONs or the MAX_NUM_THREADS. I.e. adding another SECTION will create two threads, having 5 sections will result in 4 threads as NT==4. This is true for both, 4.2 and latest svn (20070622). Adding Jakub as CC.
There are several issues, will fix them on Monday.
http://gcc.gnu.org/ml/gcc-patches/2007-06/msg01867.html http://gcc.gnu.org/ml/gcc-patches/2007-06/msg01875.html
Subject: Bug 32468 Author: jakub Date: Mon Jul 2 19:19:28 2007 New Revision: 126224 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=126224 Log: PR libgomp/32468 * sections.c (GOMP_parallel_sections_start): Only decrease number of threads to COUNT if dyn_var is true. * testsuite/libgomp.c/pr32468.c: New test. Added: trunk/libgomp/testsuite/libgomp.c/pr32468.c Modified: trunk/libgomp/ChangeLog trunk/libgomp/sections.c
Subject: Bug 32468 Author: jakub Date: Mon Jul 2 19:22:47 2007 New Revision: 126226 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=126226 Log: PR libgomp/32468 * omp-low.c (check_combined_parallel): New function. (lower_omp_parallel): Call it via walk_stmts, set OMP_PARALLEL_COMBINED if appropriate. (determine_parallel_type): If OMP_FOR resp. OMP_SECTIONS isn't the only statement in WS_ENTRY_BB or OMP_RETURN the only one in PAR_EXIT_BB and not OMP_PARALLEL_COMBINED, don't consider it as combined parallel. * gcc.dg/gomp/pr32468-1.c: New test. Added: trunk/gcc/testsuite/gcc.dg/gomp/pr32468-1.c Modified: trunk/gcc/ChangeLog trunk/gcc/omp-low.c trunk/gcc/testsuite/ChangeLog
Subject: Bug 32468 Author: jakub Date: Mon Jul 2 19:26:25 2007 New Revision: 126227 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=126227 Log: PR libgomp/32468 * sections.c (GOMP_parallel_sections_start): Only decrease number of threads to COUNT if dyn_var is true. * testsuite/libgomp.c/pr32468.c: New test. Added: branches/gcc-4_2-branch/libgomp/testsuite/libgomp.c/pr32468.c Modified: branches/gcc-4_2-branch/libgomp/ChangeLog branches/gcc-4_2-branch/libgomp/sections.c
Subject: Bug 32468 Author: jakub Date: Mon Jul 2 19:27:28 2007 New Revision: 126228 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=126228 Log: PR libgomp/32468 * omp-low.c (check_combined_parallel): New function. (lower_omp_parallel): Call it via walk_stmts, set OMP_PARALLEL_COMBINED if appropriate. (determine_parallel_type): If OMP_FOR resp. OMP_SECTIONS isn't the only statement in WS_ENTRY_BB or OMP_RETURN the only one in PAR_EXIT_BB and not OMP_PARALLEL_COMBINED, don't consider it as combined parallel. * gcc.dg/gomp/pr32468-1.c: New test. Added: branches/gcc-4_2-branch/gcc/testsuite/gcc.dg/gomp/pr32468-1.c Modified: branches/gcc-4_2-branch/gcc/ChangeLog branches/gcc-4_2-branch/gcc/omp-low.c branches/gcc-4_2-branch/gcc/testsuite/ChangeLog
Fixed in SVN.