We'll want to do some benchmarking of this to explore different heuristics for which allocations to merge. One interesting heuristic would be to merge allocations where the address of one allocation is stored in memory allocated by another, on the theory that when the outer one needs to be collected, the inner one will, too (though not necessarily vice versa).

Linked structures are likely a good candidate also, as having the structures allocated close by in memory could be good for cache performance.

In the general case merged allocations don't work with GCs because many GCs don't support interior pointers. That is if the GC knows it has allocated the address 0xABC0 but you have a pointer to 0xABC8, the GC might not recognise this as a pointer and might free your memory early. So you want to check if your GC supports interior pointers and what the limits are. But your specific case where A always points to B is safe if you know that A always has a longer-or-equal lifetime than B.

One optimisation some GCs support, usually those with bump-allocation for the nursery (or semispace). Is that a test-for-space can be separated from the actual allocation. If an allocation looks like this:

if (bump_pointer + size_to_allocate >= limit) {
   do_gc();
}
new_object = bump_pointer;
bump_pointer += size_to_allocate;

And this code would be inlined on every allocation site, then when there are multiple allocations you can instead put:

if (bump_pointer + total_size >= limit) {
   do_gc();
}
new_object_1 = bump_pointer;
bump_pointer += size_1;
new_object_2 = bump_pointer;
bump_pointer += size_2;
new_object_3 = bump_pointer;
bump_pointer += size_3;

There must be no calls to other functions that could allocate between the test and last allocation.

Each of these allocations is still a separate heap object so you don't save memory or other execution costs, only the test is amortised. But on the other hand the object's lifetimes can be completely unrelated.

I checked with @jimbxb, Since Wybe uses BoehmGC the "if they have related lifetimes and one always points to the other" optimisation is available, so is some level of interior pointer support, but if you want full support it may be an extra compile time option and cause the GC to retain more memory due to having more false-positive pointers.