[llvm] [InstCombine] Transform high latency, dependent FSQRT/FDIV into FMUL (PR #87474)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> @arsenm fmul(x, 1/y) is being transformed to fdiv(x, y) post my transform(in the instcombine pass itself). This is deleting some of the fastmath flags and fpmath metadata and I verified this by dumping the tests output post transform. This issue has started happening recently. So, expect some missing flags/metadata in the tests

For example, what cases? I think it's quite important to try to preserve !fpmath (and a lesser extent the flags). For AMDGPU losing these can trivially result in a net loss 

https://github.com/llvm/llvm-project/pull/87474