I wish there were a better way to test this but everything I can think of in Ruby would defeat the optimization.

Perhaps we could introduce something that tests the IR directly? Confirm that something like eval "foo rescue nil" includes the appropriate instruction to disable backtrace?

Looks good in any case, thank you!

hey, thanks for the merge and response.

had the idea of doing like a 'performance' unit test (haven't looked into it).
establish a Java stack-trace-less exception baseline cost and assert our stack-trace less pass are not order of magnitude slower.

That could work too. My main worry about performance tests is that Travis will sporadically be slow, or become slow, for a given build. It would be better than nothing though.

We could also look at just setting up a periodic dedicated-server run of all benchmarks we've committed, tracking them over time like the Pypy project has done. Then at least we could get some mild notification if something regresses seriously.