ScopeStack implementation stolen from Andreas Fredriksson's DICE presentation on Scope Stack Allocation.

ScopeStackDemo.cpp

// Simplified scope stack implementation

#include <new>
#include <cstdio>
#include <cassert>

typedef unsigned char u8;

class LinearAllocator {
public:
	static const size_t Alignment = 16;

	LinearAllocator(void *ptr_, size_t size)
	:	m_ptr((u8*)ptr_)
	,	m_end(m_ptr + size)
	{}

	static unsigned int alignedSize(size_t size) {
		return (size + (Alignment - 1)) & ~(Alignment - 1);
	}

	void* allocate(size_t size) {
		size = alignedSize(size);
		u8 *result = m_ptr += size;
		assert(result + size <= m_end);
		return result;
	}

	void *ptr() { return m_ptr; }

	void rewind(void *ptr) {
		m_ptr = (u8*) ptr;
	}

private:
	u8 *m_ptr, *m_end;
};

struct Finalizer {
	void (*fn)(void *ptr);
	Finalizer *chain;
};

template <typename T>
void destructorCall(void *ptr) {
	static_cast<T*>(ptr)->~T();
}

class ScopeStack {
private:
	LinearAllocator& m_alloc;
	void *m_rewindPoint;
	Finalizer *m_finalizerChain;

	static void* objectFromFinalizer(Finalizer *f) {
		return ((u8*)f) + LinearAllocator::alignedSize(sizeof(Finalizer));
	}

	Finalizer *allocWithFinalizer(size_t size) {
		return (Finalizer*) m_alloc.allocate(size + LinearAllocator::alignedSize(sizeof(Finalizer)));
	}

public:
	explicit ScopeStack(LinearAllocator& a)
	:	m_alloc(a)
	,	m_rewindPoint(a.ptr())
	,	m_finalizerChain(0)
	{}

	~ScopeStack() {
		for (Finalizer *f = m_finalizerChain; f; f = f->chain) {
			(*f->fn)(objectFromFinalizer(f));
		}
		m_alloc.rewind(m_rewindPoint);
	}

	template <typename T>
	T* newObject() {
		// Allocate memory for finalizer + object.
		Finalizer* f = allocWithFinalizer(sizeof(T));

		// Placement construct object in space after finalizer. Do this before
		// linking in the finalizer for this object so nested calls will be
		// finalized after this object.
		T* result = new (objectFromFinalizer(f)) T;

		// Link this finalizer onto the chain.
		f->fn = &destructorCall<T>;
		f->chain = m_finalizerChain;
		m_finalizerChain = f;
		return result;
   	}

	template <typename T>
	T* newPOD() {
		return new (m_alloc.allocate(sizeof(T))) T;
   	}

	// C++ suckage: need overloads ov newObject() for many arguments..
};

struct Foo {
	static int count;
	int num;
	Foo() : num(count++) { printf("Foo ctor %d\n", num); }
	~Foo() { printf("Foo dtor %d\n", num); }
};
int Foo::count;

u8 test_mem[65536];

int main()
{
	LinearAllocator allocator(test_mem, sizeof(test_mem));

	ScopeStack outerScope(allocator);
	Foo* foo0 = outerScope.newObject<Foo>();

	{
		ScopeStack innerScope(allocator);
		Foo* foo1 = innerScope.newObject<Foo>();
		Foo* foo2 = innerScope.newObject<Foo>();
	}

	return 0;
}

Unless I am somehow mistaken line 24 is incorrect. m_ptr is incremented by size before assigning to result on my machine. wouldn't you want to return the current m_ptr location and then increment m_ptr by size? The slides showed this too and I didn't understand it, maybe it was just an error. Your test wouldn't show the bug because all your objects are the same size, but if they were different sizes I think you'd have memory corruption.