32bits shd mem + Unordered crash in rb_tree when clearing in /O2 #75
Description
I've set up a Boost IPC shared memory for 32/64bits communication as per comment from Boost header so offset_ptr<SharedType, int32_t, uint64_t, sizeof(std::uint64_t)> along with rbtree_best_fit's MemoryAlignment sizeof(std::uint64_t)
Boost 1.68
In x86 VS 15.8.8 with /O2 + /Ob2 + /Oi + /Ot + /Oy and BOOST_INTERPROCESS_DISABLE_FORCEINLINE
The rbtree_algorithms will crash trying to rebalance after erasure. But if I disable the optimisation /Od, everything will be fine.
This is my example program:
template<typename SharedType>
using OffsetPtr = boost::interprocess::offset_ptr<SharedType, int32_t, uint64_t, sizeof(std::uint64_t)>;
using ManagedWindowsSharedMemory =
boost::interprocess::basic_managed_windows_shared_memory<char,
boost::interprocess::rbtree_best_fit<boost::interprocess::mutex_family, OffsetPtr<void>, sizeof(std::uint64_t)>,
boost::interprocess::iset_index
>;
using ShMemSegmentManager = ManagedWindowsSharedMemory::segment_manager;
using shmMultiMapAllocator = boost::interprocess::allocator<std::pair<unsigned int const, unsigned int>, ShMemSegmentManager>;
using shmMultiMap = boost::unordered_multimap<unsigned int, unsigned int, boost::hash<unsigned int>, std::equal_to<unsigned int>, shmMultiMapAllocator>;
// Win32: /O2 + /Ob2 + /Oi + /Ot + /Oy + BOOST_INTERPROCESS_DISABLE_FORCEINLINE
// crash at rbtree_algorithms::rebalance_after_erasure_restore_invariants
// const node_ptr x_parent_left(NodeTraits::get_left(x_parent));
// 0xC0000005: Access violation reading location 0x00000008.
int main(int argc, char **argv)
{
auto managedShMem = new ManagedWindowsSharedMemory(boost::interprocess::create_only, "shmName", 2*1024*1024, 0);
auto mSpecimenToFileMap = managedShMem->find_or_construct<shmMultiMap>("specimenToFileMap")
(3, boost::hash<unsigned int>(), std::equal_to<unsigned int>(), shmMultiMapAllocator(managedShMem->get_segment_manager()));
for (int i = 0; i < 5000; ++i)
{
mSpecimenToFileMap->insert(std::make_pair<unsigned int, unsigned int>(std::move(unsigned int(i/10)), std::move(unsigned int(i))));
}
mSpecimenToFileMap->clear();
system("pause");
return 0;
}
In x64 without BOOST_INTERPROCESS_DISABLE_FORCEINLINE, I do not have any issue.
In x86, if I use simple_seq_fit instead of rbtree_best_fit, it will not crash with the optimisation on.