I have they are rare though. once ever 10 years for a good sized project seems right. Where they habpen I wouldn't expect anything other than the exact hardware to reproduce them so I'm still not wanting a workstation live the target.
I've run into tons of optimizer bugs. While it's possible they would require the exact hardware to reproduce them, that has never been the case. Generally you just need something to run the actual output of the optimizer. You can even reproduce the bugs in emulators, it just takes way longer.
Most optimizer bugs are about the optimizer making assumptions about properties of the source code, not assumptions about properties of the hardware. Certainly it can go both ways, but of course you want to test the actual output of the optimizer. Doing otherwise is not at all what I'd call standard practice.
