Renaming allows to to eliminate write after read dependencies. E.g. Computing a value, storing it to AX, doing something with AX, computing another value, storing it to AX, doing something with AX. If the two computations are independent, the AX dependency is false and the OOO core can execute them concurrently by storing the results into two different rename registers and rewriting the "doing something with AX" to source their inputs from those two different registers instead.

In terms of live ranges, the first value stored in AX has a live range that does not overlap that if the second value stored in AX. So the compiler can reuse AX for those two values without problems.

That's not true in the scenario of the article without moving code around. The live ranges overlap--you have to keep around two or more distinct values before any of them are used. You can't store the results into the same register without clobbering one of the values. So the compiler is forced to introduce a memory operation by spilling a value, and the CPU can't do anything to eliminate that after the fact.