[compiler-rt] Bfi precision (PR #66285)

[Nikita Popov via llvm-commits]

nikic wrote:

> Oh so this isn't even a PGO enabled build, interesting...

Ah yes, I should have mentioned that. This is indeed without PGO. So I expect this is related to the impact of the estimated BFI on the inlining cost model (which is the only significant place using BFI in non-PGO builds). (I've seen issues with bad inlining decisions due to BF underflow in the past, not sure whether this change makes that better or worse.)

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