[llvm] Bfi precision (PR #66285)

[via llvm-commits]

qcolombet wrote:

FWIW, I'm supportive of the investigation here.

Just one comment:
> I am still collecting experiences here, but my gut feeling would be that we rather want to differentiate the big values than the small ones...

I think so, but we probably will lose some valuable information too if we collapse small values.
I expect this will affect hot/cold placement if e.g., two cold blocks have the same weight now and ditto for spill placement in regalloc. I guess it really depends how we "collapse" frequencies (e.g., do we lose 5% accuracy or do we lose 2x for these). In any case, the cases where this matters and overflow occurs are hopefully going to be rare.

Long story short, we'll have regressions :).

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