Write swap in assembly instead of C++

[eholk]

Right now, C++ compilers like to make set_registers into a tail call, which causes problems when get_registers returns the second time. To get around this, we should write the swap context function in assembly instead of disabling compiler optimizations.

[marijnh]

Could you write it in LLVM bitcode instead? I think we're hoping to use as little assembly as possible, in order to make supporting different processor types easy.

[eholk]

My first thought was no, because swapping contexts needs direct access to the machine's registers. However, we already have set_registers and get_registers written in assembly, so we could probably write the swap function in LLVM.

If we do this, can we guarantee that LLVM won't turn the call to set_registers into a tail call like is happening in the C version?