Earlier quoted context omitted.
Another approach is to do the opposite - always store all local variables in the heap, and rely on scalar replacement of aggregates to put them back onto the stack where possible. This is what I do in my Ruby interpreter. That’s definitely not compatible with a single-pass though!
A third option that can still be single pass is the Smalltalk way, where the stack frames themselves are first-class objects: the closures then keep a reference to the stack frame and the GC won't collect any live stack frames, which will contain closed-over variables.
Yes that's really what I'm talking about, but plus the (really required for practical performance) optimisation of virtualising the frames onto the stack if they don't escape an activation.