The point of an optimizing compiler is that it should generate an optimized program that is functionally equivalent to the original. Functional equivalent has to do with observable effects, of course; the state would naturally not be the same. Any compiler that can generate an incorrect translation is bad, in my opinion, although I suppose that you could relax certain issues in different situations. One advantage of a mainstream compiler is that bugs are fixed faster.
As for performance, compilers generally do two types of optimizations: platform-independent optimizations (at the source level/syntax tree/intermediate representation), and platform-specific ones.
I would guess that the best platform-independent optimizations are probably in various academic modifications of existing compilers (but probably a different optimization and each). These optimizations are eventually recognized and eventually trickle into the mainstream.
As for platform-specific optimizations, I think that companies making these platforms often have the best optimizations (e.g., the internal Intel compilers), but they are not yet integrated into general-availability compilers, and eventually trickle down as well.
