IR2Native Model? about ml-compiler-opt HOT 1 CLOSED

Sorry for the delay in replying.

The code extracting the data for IR2Native would have been in llvm - it's essentially just extracting the features used llvm/lib/Analysis/InlineSizeEstimatorAnalysis.cpp, right after inlining (since then the function's body is all that's necessary to predict its final size - i.e. you don't need to look at other functions), and then the size expression during AsmPrint - ing, and dumping them out.

We're not using it anymore though - the current training algorithms use final size instead (at a point, we should probably yank out InlineSizeEstimatorAnalysis, too)

If you're interested in size estimation with supervised learning, please add it to the agenda for the next LLVM MLGO meeting - I believe there are some folks who may have some ideas, it'd be an interesting discussion!

Related, we had a while back an rfc about whether we'd benefit from some facility for embedding data in a section in the binary - this would have been one of the motivating scenarios - if the need resurfaces, I'm sure there'll be other benefiting scenarios (there some interest on the thread, IIRC)

from ml-compiler-opt.