Implementing an overflow-free system stack in C90

See GCC's split-stack capability. I believe this was originally implemented to support Go. It pretty much works as you suggest.

EDIT: The commentary below discusses another system that does heap-allocation of activation records.

You can do this - I believe modern gcc may even have an option for it - but it greatly increases the cost of function calls and has very little practical benefit. Especially on modern systems with 64-bit addressing, there's plenty of address space for each thread to have its own stack plenty far from every other thread's stack. If you find yourself using more than logarithmic-scale call recursion, something is wrong with your algorithms anyway...