Closed Quincunx271 closed 4 years ago
This bug also happens with the PRP spill cost function so there may be something wrong with history domination.
Probably has something to do with a898f8d75c74861888f5f3f3dc39dc068a89dfd6
You could try manually reverting that change. It would be nice to have compiler temps so that we can add a testcase for history domination once this is fixed. Also I wont be able to debug this myself without the IR.
Probably can't give compiler temps because of licensing for SPEC CPU2006. It may be worth replicating this with our own code.
When running cactus with PERP and APPLY_HISTORY_DOMINATION set to YES, the
ApplyBndRobin:63
block has a spill cost of 4, when it should be 2.sched.ini file follows. I set
ApplyBndRobin
as the only hot function so that it wouldn't waste time compiling the rest of the benchmark:The relevant part of the log (actual):
The relevant part of the log after history domination is turned off (expected):