Optimization question

[James Demmel]

Since floating point optimizations are obviously very useful to a large
number of users much of the time, and poisonous to a few important codes
written by experts, why can't we have both worlds via a pragma or other
construct which can turn off optimization for selected stretches of code
so that "what you see in the source is what you get"? The experts who write
this stuff would have no difficulty deciding where to insert these pragmas,
and the average user who doesn't care gets optimization by default. I
suppose one could also do it via a compiler flag, to turn off all
optimization in a subroutine, but that is clearly less convenient.

What are the objections, which I may not see, to this approach?
  
   Jim Demmel