[llvm] Bfi precision (PR #66285)

[via llvm-commits]

WenleiHe wrote:

> This is concerning. Can this be reverted for now and we can help with some internal performance testing. @xur-llvm

Internally, with this change, on a few large workloads, we saw ~0.2-0.5% performance improvements, all with 2-3% .text size reduction at the same time. Our testing covered both sampling pgo and instrumentation pgo.

The change makes profile more accurate, which is a good improvement by itself. But PGO performance isn't just about profile accuracy, and it depends on training quality and representativeness too. Hypothetically, if some of the training isn't that representative relative to perf testing scenario, by making profile more accurate, we could also make profile deviate more from what's needed for best perf on testing scenario. 

That said, I think it'd be fair to revert if this is proven to be a general regression. For that, can we wait for your results before reverting? 

https://github.com/llvm/llvm-project/pull/66285