Example where bugpoint does a better job than llvm-reduce

[arsenm]

sample.zip
Here is a sample reduced testcase where bugpoint was able to improve upon llvm-reduce's result. After feeding llvm-reduce's output into bugpoint, it cut out a few hundred more lines of IR. It may or may not be useful to look at the diff to figure out what bugpoint is doing better.

The exact reproduction steps are hard to share. The original reproducer is a register allocation issue, which are always super sensitive to the exact code. So far I've only been able to reproduce it using a specific clang build (the one provided by the rocm 5.0 release version)

[regehr]

I don't think I'm going to try to reproduce this, but I'm very interested in specific things to add to llvm-reduce to help it get unstuck.

but also I just pushed a change to llvm-reduce that makes it run to fixpoint instead of running through its set of passes just once. unless you were previously using the -max-pass-iterations command line option, this new change might improve its output. if you don't mind building a fresh llvm and running llvm-reduce on your example again, I'd be interested to hear about the results

[arsenm]

I don't see anything actionable here without a reproducer script. I do think llvm-reduce can cleanup some of the remaining things in this sample that it couldn't before.