eon performance regression

[Jason Merrill]

It seems possible that my patch is responsible, perhaps because of code
growth due to calling the destructors at each call site rather than at a
single place in the callee, but I still wouldn't expect it.  It's really
impossible to do more than speculate wildly without access to the code
itself; can you see a significant difference in the generated code?  Can
you run it with profiling?  Can you try running the benchmark without my
patch?  It seems to me that more speculation is not very useful at this
point.

Jason