[PATCH] Fix inliner to not inline functions that call setjmp() using "invoke"
chandlerc at gmail.com
Thu Dec 12 23:48:21 PST 2013
This is horrible. Horrible. Thank you for uncovering this.
You are on the right track with just visiting the terminator. It was never a good idea to not do this. It was a terrible, terrible idea. (Also, my idea. Doh!)
I've got a patch series to fix all of this (thanks for the test case btw!). However, the initial change to start actually accounting for the cost of invokes and other terminators has the potential to *radically* change our performance numbers. I ran a bunch of benchmarks overnight last night (sorry for not sending this out) and everything looks good so I'm about to commit.
Again, thanks for the investigative work here. I really don't know how any of this worked....
More information about the llvm-commits