include/parallel/compatibility.h uses an OpenMP critical section on all platforms except x86_64, which issues messages that break the check-parallel testsuite: Running /home/jwakely/src/gcc/libstdc++-v3/testsuite/libstdc++-dg/conformance.exp ... ERROR: tcl error sourcing /home/jwakely/src/gcc/libstdc++-v3/testsuite/libstdc++-dg/conformance.exp. ERROR: could not compile testsuite_abi.cc while executing "error "could not compile $f"" (procedure "v3-build_support" line 61) invoked from within "v3-build_support" (file "/home/jwakely/src/gcc/libstdc++-v3/testsuite/libstdc++-dg/conformance.exp" line 25) invoked from within "source /home/jwakely/src/gcc/libstdc++-v3/testsuite/libstdc++-dg/conformance.exp" ("uplevel" body line 1) invoked from within "uplevel #0 source /home/jwakely/src/gcc/libstdc++-v3/testsuite/libstdc++-dg/conformance.exp" invoked from within "catch "uplevel #0 source $test_file_name"" testsuite/parallel/libstdc++.log shows: In file included from /home/jwakely/build/i686-pc-linux-gnu/libstdc++-v3/include/parallel/find.h:40:0, from /home/jwakely/build/i686-pc-linux-gnu/libstdc++-v3/include/parallel/algobase.h:42, from /home/jwakely/build/i686-pc-linux-gnu/libstdc++-v3/include/bits/stl_algobase.h:1222, from /home/jwakely/build/i686-pc-linux-gnu/libstdc++-v3/include/algorithm:62, from /home/jwakely/build/i686-pc-linux-gnu/libstdc++-v3/include/i686-pc-linux-gnu/bits/stdc++.h:65, from <command-line>:0: /home/jwakely/build/i686-pc-linux-gnu/libstdc++-v3/include/parallel/compatibility.h: In function 'int64_t __gnu_parallel::__fetch_and_add_64(volatile int64_t*, int64_t)': /home/jwakely/build/i686-pc-linux-gnu/libstdc++-v3/include/parallel/compatibility.h:167:42: note: #pragma message: slow __fetch_and_add_64 #pragma message("slow __fetch_and_add_64") ^ /home/jwakely/build/i686-pc-linux-gnu/libstdc++-v3/include/parallel/compatibility.h: In function 'bool __gnu_parallel::__compare_and_swap_64(volatile int64_t*, int64_t, int64_t)': /home/jwakely/build/i686-pc-linux-gnu/libstdc++-v3/include/parallel/compatibility.h:320:45: note: #pragma message: slow __compare_and_swap_64 #pragma message("slow __compare_and_swap_64") ^ The parallel/compatibility.h header checks for #if defined(__x86_64) [use atomics] #elif defined(__i386) && \ (defined(__i686) || defined(__pentium4) || defined(__athlon) \ || defined(__k8) || defined(__core2)) [use atomics] #else [use omp critical] #endif But __i686 isn't defined on i686, for example, and this doesn't handle platforms such as power64 which support the necessary 64-bit ops. Instead of a (broken) whitelist of targets the code could use __atomic_always_lock_free to check for the availability of atomic ops on the target. Another possibility would be to add a new configure option: --enable-libstdcxx-parallel-atomics=[omp|libatomic] The default would be to use the omp criticial section as it is now, the libatomic option would force use of atomics, which would require libatomic for platforms where the built-ins call a library function.
Created attachment 28301 [details] Use built-in atomics on non-x86_64 targets. This patch uses __atomic_always_lock_free to check if the __atomic_* built-ins can be used, which allows them to be used on more platforms than the current preprocessor checks allow.
Jon, please, pursue the approach you like better and let's make progress on this: remember that we are still in Stage 1, we can also take some little risks.
Author: redi Date: Tue Oct 9 08:16:13 2012 New Revision: 192240 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=192240 Log: PR libstdc++/54754 * include/parallel/compatibility.h: Use atomic built-ins when they are lock-free. Modified: trunk/libstdc++-v3/ChangeLog trunk/libstdc++-v3/include/parallel/compatibility.h
fixed