But the real issue is that what he's doing is completely unnecessary, because you could manage multiple stacks yourself using ordinary C data structures and ordinary C code. Evidently he thinks he's gaining some kind of speed advantage by doing things this way, but the risk is very high versus the gain. This kind of optimization may have worked OK on a DOS machine, but it's neither safe nor obviously faster to do things like this on a modern CPU and OS.
In this case I think the harsh criticism was fully justified. People should be warned away from doing this sort of thing, precisely because it's hard to get right. As the thread makes abundantly clear the original programmer was fairly clueless, otherwise he wouldn't have gone off half-cocked thinking that the fault was with cygwin or somesuch. He was making his own dynamite and was too clueless to realize what he was doing wrong.
Also depending on how the compiler sets up the runtime environment, setting ESP is not sufficient. Intel also has the segment registers. EBP is the stack segment pointer. I'm pretty sure the heap is not in the EBP segment.