[llvm] Bfi precision (PR #66285)

[Matthias Braun via llvm-commits]

MatzeB wrote:

Thanks for confirming @duncan!

The current state is mostly a result of http://reviews.llvm.org/D8718 it seems and I am trying to understand the intentions there or whether that could have been an accident.

I am definitely running into overflow and noise problems with this patch applied and am working on fixing these issues as I see them, hopefully this is manageable.

Another discussion here is whether to normalize "min" value to integer 1 and have higher precisions for the small numbres while saturating big numbers. Or normalize "max" to `UINT64_MAX` which I do here, which may map multiple smaller values to integer 1. 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...

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