[llvm] [GlobalISel] Lower optimized sqrt libcalls to G_FSQRT (PR #221747)

[Kiran Chandramohan via llvm-commits]

kiranchandramohan wrote:

> Should not do this. We're trying to remove this code in selectiondag. There's no reason SimplifyLibcalls can't change libcalls to intrinsics

Thanks, yes if it can be done earlier in SimplifyLibCalls that would be better. Is there a PR in progress? 

For context, this fixes a performance diff with SelectionDag in Spec2017 544.nab_r at `O3`.

https://github.com/llvm/llvm-project/pull/221747