Upstream notes on version 6.4:

This is a minor new release, the API and ABI (Application binary interface) are backwards compatible.

This version fixes a critical integer overflow in the B formula which is part of Xavier Gourdon's prime counting algorithm. The integer overflow occurred when computing B(x) with x > 1e25, this bug has been present in primecount-6.1, 6.2 and 6.3. This version also adds a workaround for a severe scaling issue in Clang's OpenMP library (Clang 11, 2021) that occurs when using dynamic thread scheduling combined with long running computations. Because of this issue primecount-6.3 compiled with Clang takes 2.5x longer to compute AC(1e24) than primecount-6.3 compiled with GCC on my dual-socket AMD EPYC Rome server. I have submitted a bug report to LLVM which contains more information: https://bugs.llvm.org/show_bug.cgi?id=49588

• B.cpp: Fix integer overflow (thanks to David Baugh for reporting it)
• B_mpi.cpp: Fix integer overflow (same as in B.cpp).
• P2.cpp: Fix integer overflow (same as in B.cpp).
• for_atomic.hpp: Lock-free for loop built with std::atomic.
• AC.cpp: Fix Clang/OpenMP scaling issue.
• AC.cpp: Decrease size of SegmentedPiTable by 1.5x.
• AC_libdivide.cpp: Fix Clang/OpenMP scaling issue.
• AC_libdivide.cpp: Decrease size of SegmentedPiTable by 1.5x.
• S2_easy.cpp: Fix Clang/OpenMP scaling issue.
• S2_easy_libdivide.cpp: Fix Clang/OpenMP scaling issue.
• SegmentedPiTable.cpp: Increase minimum segment size.
• SegmentedPiTable.cpp: Annotate pi(x) with ALWAYS_INLINE.
• Sieve.cpp: Simplify counters_dist initialization.
• PiTable.cpp: Annotate pi(x) with ALWAYS_INLINE.
• LoadBalancerP2.cpp: Fix last iteration detection.