I should mention that I have the Intel C++ compiler (trial) for MacOS. It's not at all reliable and I can get it to crash with some of the crappy code at work! It has picked up some genuine bugs though!
However, I'll compile a Quake 3 with it to compare the performance. The damn thing will parallelize and vectorize your code for you for free!!! I'm excited about that :-D
I compiled some bundles at work which're used by ColorTuneX (if you know what I mean) and it shaved a second off generating an ICC device link profile... i.e. 23 seconds to 22 (not much!). I then profiled it with Shark and found that 25% of that time measurement was then in GCC compiled code which was only optimized by GCC not ICC. It didn't used to spend it's time there from looking at a GCC only build.
I tried to compile that bundle but the icc compiler crashed compiling some ropey code (not mine). I'll speak to the offending team (Windows) to see if I can force them to make that code simpler so that icc can handle it.
At 600 bucks it seems to be worth it *but* you have to have respectable code to pass through it. I did file a defect with Intel for not supporting @executable_path (frameworks and dylibs) but not the compiler crash.