The optimizer assumes that a non-atomic non-volatile value written to memory stays the same, until some code that could modify it is executed. This allows a lot of obvious optimizations, like hoisting needlessly repeated computation out of loops, removal of redundant loads, and optimizations of common subexpressions and arithmetic.

If a value in memory could suddenly revert to something else, then

   if obj.x == 1 {
       print(obj.x)
   }

could print "2", and such paradoxes can lead to unsafety:

   if obj.x < array.len {
       array[obj.x]
   }

Defining that values in memory can't be trusted would mean giving up on a lot of optimizations, and require implementations to emit a lot of mostly-useless copying of values to protect them from being unexpectedly modified.