Not an expert in the field, but it's possible to convolve a signal on the recorded impulse response of a room [1] or a 'Slinky' toy [2], or that of a subjectively 'desirable' piece of audio equipment [3].
The result is as if your signal were played in that room ( with all its reverb ) or through that thing ( with its 'desirable'-or-otherwise characteristics modeled in its impulse response ).
Thinkable in 1962, but not thought possible. Now it's part of an industry.
To add an aside to my other comment, as a grad student in a spatial audio class, I led a project to create a new method of simulating physical reverb by using mics and speakers pointed at each other in an anechoic space, letting the mutual feedback create reverb. You could apply delay to "push back" the walls and apply filters matching the absorption characteristics of different materials. It was a fun project, which won my group Grad Project of the Year at my school's demo day. A couple commercial products and systems use similar concepts to achieve active acoustic control.
The downside of convolutional reverb is the lack of parameterization. You're kind of stuck with one fixed geometry of source, receiver and surfaces. It can also be expensive to apply in real-time processing.
A lot can be done post-processing the impulse response (volume envelopes, timestretching, combining with other parts, etc)
As for the efficiency: a modern laptop can easily run ~100 channels of multi-second convolution reverb in realtime on 44.1/48kHz sample rate and <10ms latency in real-time on 1 core.
Yeah, but if you post-process the impulse response, it's difficult to end up with something else that "looks" as much like a real impulse response. If that's important to you.
Don't get me wrong, I'm not trying to argue against using convolutional reverb. I was just throwing out a couple reasons why people still use other approaches.
The paper is unfortunately behind a pay wall: http://www.aes.org/e-lib/browse.cfm?elib=16101