JamesMunns[m]: > <@jamesmunns:beeper.com> I feel like this should work, but I can't tell you why it doesn't.
> 
> have you guaranteed the kernel is actually doing good ops to the right memory, and it's not a bug somewhere else? and peeked at the asm to see if it is just doing garbage (and maybe diff with compiler fence vs the fence?)

"guaranteed" well. Stared at both code and disassembly, and stepped through the fully LTO'ed code a lot, to the point where I can say that the "write the address" part has been executed according to gdb's "x/x <addr>".