[Bug c/98965] assignment to a struct with an atomic member not atomic
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98965 --- Comment #5 from Martin Sebor --- I was thinking of the latter, i.e., splitting the assignment into a series of ordinary and atomic, the latter per member. The union case would be diagnosed (it too should be straightforward to detect). But C++ rejects such mixed assignments, so the easiest solution would be to do the same in C (or at least warn).
[Bug c/98965] assignment to a struct with an atomic member not atomic
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98965 --- Comment #4 from joseph at codesourcery dot com --- The difficulty with making such an assignment atomic is that atomic operations for different sizes of atomic access don't interoperate on the same memory; if the struct contains an _Atomic int, that's accessed atomically via appropriate int-sized atomic operations, while a larger structure may use locking for atomic access in libatomic, and if both mechanisms are used on the same memory the result doesn't properly follow the memory model. So you'd need to split struct assignment up into assignment of members to access a member atomically. But that couldn't work at all when you have a union with atomic members, because in the union case the compiler can't know which is the currently active member that would determine the size of the atomic access.
[Bug c/98965] assignment to a struct with an atomic member not atomic
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98965 Martin Sebor changed: What|Removed |Added Keywords||diagnostic --- Comment #3 from Martin Sebor --- The topic came up in a discussion on the WG14 reflector. For waht it's worth, my impression is that it's simply an oversight, a result of an incomplete integration of atomics into the core language. It would be hard to justify making this (deliberately) undefined. The member type information is readily available at the point of the assignment so if it's not meant to be supported the standard could easily require a diagnostic without imposing a great burden on implementations. Until or unless the standard is clarified, unless we choose to implement the assignment atomically, issuing a warning would help prevent bugs. As a data point, of the compilers on Godbolt only Clang and GCC support _Atomic structs and neither locks the member.
