Dmitry, You've introduced atomic_uint64_t stats_[AllocatorStatCount]; in http://llvm.org/viewvc/llvm-project?view=revision&revision=173332 Do you mind to change it to atomic_uintptr_t? There is of course a chance of overflow on 32-bit arch (the number of mallocs in a process may easily go over 2^32 in a long run), but this is just stats, I think we can tolerate it.
--kcc On Wed, May 28, 2014 at 2:41 AM, Ramana Radhakrishnan <ramana....@googlemail.com> wrote: > On Tue, May 27, 2014 at 11:00 PM, Jakub Jelinek <ja...@redhat.com> wrote: >> On Tue, May 27, 2014 at 11:50:33PM +0200, Jakub Jelinek wrote: >>> On Tue, May 27, 2014 at 04:02:08PM -0500, Peter Bergner wrote: >>> > It's being called form basically two files: >>> > >>> > [bergner@makalu-lp1 gcc-fsf-mainline-asan-debug]$ find . -name '*.o' | >>> > xargs nm -AC | grep sync_fetch_and_add_8 >>> > ./powerpc64-linux/32/libsanitizer/sanitizer_common/.libs/sanitizer_allocator.o: >>> > U __sync_fetch_and_add_8 >>> > ./powerpc64-linux/32/libsanitizer/sanitizer_common/sanitizer_allocator.o: >>> > U __sync_fetch_and_add_8 >>> > ./powerpc64-linux/32/libsanitizer/asan/.libs/asan_allocator2.o: U >>> > __sync_fetch_and_add_8 >>> > ./powerpc64-linux/32/libsanitizer/asan/asan_allocator2.o: U >>> > __sync_fetch_and_add_8 >>> >>> Does ppc32 have any atomic 64-bit loads/stores (in the sense that the >>> aligned >>> 64 bits are written as one memory transaction, not each 32-bit word >>> separately)? >>> In any case, atomic_uint64_t there seems to be used only for some statistic >>> counter and not really atomic anyway, as additions aren't performed using >>> atomic instructions, but just atomic load, followed by normal arithmetics, >>> followed by atomic store. >>> Can't 32-bit counters be used instead on 32-bit arches? >>> >>> I see there is another spot with atomic_uint64_t in sanitizer_lfstack.h, >>> but for some reason it isn't used now at all (there it would want to use >>> 64-bit compare and exchange). >> >> On ARM, while it supposedly links, because __sync_compare_and_exchange_8 >> is defined in libgcc.a, it will only work with post-2011 kernels and is >> going to be very slow (because you do a separate compare and exchange > > FTR, this call down to the library function should only be for legacy > architectures. > > On ARM we have a 64 bit atomic compare and exchange which can be done > with the ldrexd / strexd instructions at the right architecture level > (v6k and above IIRC). > > Ramana