InstrDesc contains 4 vectors (as well as some other data), so it's
expensive to copy.
Details
Details
Diff Detail
Diff Detail
- Repository
- rL LLVM
Event Timeline
Comment Actions
@orodley No problem. I don't know if you have commit access (your account looks pretty new), but I'm happy to land the patch for you later if you wish.
Comment Actions
@mattd Dean's in my time zone so I asked him before I saw your comment, but thanks for the offer.
Comment Actions
Awesome. Thanks for looking into this! I'm curious, did you discover this through some performance measuring, or did it just catch your eye when reading over the source?
Comment Actions
Yep, it was through profiling. I've been working on a new scheduler using MCA, as Clement alluded to in https://bugs.llvm.org/show_bug.cgi?id=37696.
mca::Pipeline::run ends up being the hot part, and this was a simple, easily upstreamable 5% improvement within that :)