If your dynamic allocator is a pointer-increment (as it is in most copying generational GCs), then if you have a single global nursery, then two threads allocating after each other would get adjacent locations in memory when allocating.
On SBCL, it's pretty normal to dynamically allocate a cons-cell, which is the size of two pointers. The garbage collector pays zero cost for any garbage in the nursery when it is run, so the performance overhead versus stack allocation is approximately zero.
I've seen language implementations where there is no stack at all (it's one of the most straightforward ways to implement scheme). Chicken scheme works this way, for example (it cleverly uses the C stack as the nursery, since functions never return in chicken and since a copying nursery allocation is just a pointer-increment, which is what a C stack allocation is).
On SBCL, it's pretty normal to dynamically allocate a cons-cell, which is the size of two pointers. The garbage collector pays zero cost for any garbage in the nursery when it is run, so the performance overhead versus stack allocation is approximately zero.
I've seen language implementations where there is no stack at all (it's one of the most straightforward ways to implement scheme). Chicken scheme works this way, for example (it cleverly uses the C stack as the nursery, since functions never return in chicken and since a copying nursery allocation is just a pointer-increment, which is what a C stack allocation is).