Open PandaBlueYellow opened 1 year ago
Regression on SPEC 2017 integer benchmarks has also been verified by our team. It seems to be related to basic block alignment and DSB performance and we have plans to address it.
Look forward to it! Do you have a rough estimate how soon this will be?
Hopefully we start looking at it early next year.
One more question, is propeller's goal to be always enabled like thinlto? Or do you see some restrictions that propeller would be useful case by case?
Sorry for the late reply. Propeller has a build cost (extra round of profiling and build). So the improvements must be evaluated in each case to see if they outweigh the cost.
Is there an update on this?
By the way, what is the current status of Propeller? Is it available outside of Google to download and experiment with it? Especially in comparing to BOLT.
Is there an update on this?
I'll let @rlavaee answer this question. Since he has been making improvements for propeller over the past few months.
By the way, what is the current status of Propeller? Propeller is currently actively maintained in Google, and the github.com/google/autofdo is also maintained so that it builds and works with current head llvm source tree. Although it may lack some of the newest features.
We also have a plan to move create_llvm_prof's propeller bits to llvm source tree, @rlavaee will be working on this.
@zamazan4ik We haven't looked at the DSB problem yet. We will be putting more effort into upstreaming create_llvm_prof this year so it can be conveniently used by everyone.
I tested autofdo+thinlto+propeller vs autofdo+thinlto on spec 2017. autofdo+thinlto+propeller is ~2% slower overall. Although it wins in 544 and 644 by >20%, it's worse in most other cases. Is that expected?